Seiberg–Witten–Floer stable homotopy type of three-manifolds with = 0 Geometry & Topology

advertisement
889
ISSN 1364-0380 (on line) 1465-3060 (printed)
Geometry & Topology
G
T
G G TT TG T T
G
G T
T
G
T
G
T
G
T
G
T
G
GG GG T TT
Volume 7 (2003) 889–932
Published: 10 December 2003
Seiberg–Witten–Floer stable homotopy type
of three-manifolds with b1 = 0
Ciprian Manolescu
Department of Mathematics, Harvard University
1 Oxford Street, Cambridge, MA 02138, USA
Email: manolesc@fas.harvard.edu
Abstract
Using Furuta’s idea of finite dimensional approximation in Seiberg–Witten theory, we refine Seiberg–Witten Floer homology to obtain an invariant of homology 3–spheres which lives in the S 1 –equivariant graded suspension category.
In particular, this gives a construction of Seiberg–Witten Floer homology that
avoids the delicate transversality problems in the standard approach. We also
define a relative invariant of four-manifolds with boundary which generalizes
the Bauer–Furuta stable homotopy invariant of closed four-manifolds.
AMS Classification numbers
Primary: 57R58
Secondary: 57R57
Keywords: 3–manifolds, Floer homology, Seiberg–Witten equations, Bauer–
Furuta invariant, Conley index
Proposed: Tomasz Mrowka
Seconded: Dieter Kotschick, Ralph Cohen
c Geometry & Topology Publications
Received: 2 May 2002
Accepted: 5 December 2003
890
1
Ciprian Manolescu
Introduction
Given a metric and a spinc structure c on a closed, oriented three-manifold Y
with b1 (Y ) = 0, it is part of the mathematical folklore that the Seiberg–Witten
equations on R×Y should produce a version of Floer homology. Unfortunately,
a large amount of work is necessary to take care of all the technical obstacles and
to this day there are few accounts of this construction available in the literature.
One difficulty is to find appropriate perturbations in order to guarantee Morse–
Smale transversality. Another obstacle is the existence of a reducible solution.
There are two ways of taking care of the latter problem: one could either ignore
the reducible and obtain a metric dependent Floer homology, or one could do
a more involved construction, taking into account the S 1 –equivariance of the
equations, and get a metric independent equivariant Floer homology (see [18],
[20]).
In this paper we construct a pointed S 1 –space SWF(Y, c) well-defined up to stable S 1 –homotopy equivalence whose reduced equivariant homology agrees with
the equivariant Seiberg–Witten–Floer homology. For example, SWF(S 3 , c) ∼
=
S 0 . This provides a construction of a “Floer homotopy type” (as imagined by
Cohen, Jones, and Segal in [6]) in the context of Seiberg–Witten theory. It
turns out that this new invariant is metric independent and its definition does
not require taking particular care of the reducible solution. Moreover, many
of the other complications associated with defining Floer homology, such as
finding appropriate generic perturbations, are avoided.
To be more precise, SWF(Y, c) will be an object of a category C, the S 1 –
equivariant analogue of the Spanier–Whitehead graded suspension category.
We denote an object of C by (X, m, n), where X is a pointed topological space
with an S 1 –action, m ∈ Z and n ∈ Q. The interpretation is that X has
index (m, n) in terms of suspensions by the representations R and C of S 1 .
For example, (X, m, n), (R+ ∧ X, m + 1, n), and (C+ ∧ X, m, n + 1) are all
isomorphic in C. We extend the notation (X, m, n) to denote the shift of any
X ∈ Ob C. We need to allow n to be a rational number rather than an integer
because the natural choice of n in the definition of our invariant will not always
turn out to be an integer. This small twist causes no problems in the theory.
We also use the notation Σ−E X to denote the formal desuspension of X by a
vector space E with semifree S 1 action.
The main ingredient in the construction is the idea of finite dimensional approximation, as developed by M Furuta and S Bauer in [13], [4], [5]. The
Seiberg–Witten map can be written as a sum l + c : V → V, where V =
Geometry & Topology, Volume 7 (2003)
891
Seiberg–Witten–Floer stable homotopy types
i ker d∗ ⊕ Γ(W0 ) ⊂ iΩ1 (Y ) ⊕ Γ(W0 ), l = ∗d ⊕ 6 ∂ is a linear Fredholm, self-adjoint
operator, and c is compact as a map between suitable Sobolev completions of
V. Here V is an infinite-dimensional space, but we can restrict to Vλµ , the span
of all eigenspaces of l with eigenvalues in the interval (λ, µ]. Note that λ is
usually taken to be negative and µ positive. If pµλ denotes the projection to
the finite dimensional space Vλµ , the map l + pµλ c generates an S 1 –equivariant
flow on Vλµ , with trajectories
∂
x(t) = −(l + pµλ c)x(t).
∂t
If we restrict to a sufficiently large ball, we can use a well-known invariant
associated with such flows, the Conley index Iλµ . In our case this is an element
in S 1 –equivariant pointed homotopy type, but we will often identify it with the
S 1 –space that is used to define it.
x : R → V,
In section 6 we will introduce an invariant n(Y, c, g) ∈ Q which encodes the
spectral flow of the Dirac operator. For now it suffices to know that n(Y, c, g)
depends on the Riemannian metric g on Y, but not on λ and µ. Our main
result is the following:
Theorem 1 For −λ and µ sufficiently large, the object (Σ−Vλ Iλµ , 0, n(Y, c, g))
depends only on Y and c, up to canonical isomorphism in the category C.
0
We call the isomorphism class of SWF(Y, c) = (Σ−Vλ Iλµ , 0, n(Y, c, g)) the equivariant Seiberg–Witten–Floer stable homotopy type of (Y, c).
0
It will follow from the construction that the equivariant homology of SWF
equals the Morse–Bott homology computed from the (suitably perturbed) gradient flow of the Chern–Simons–Dirac functional on a ball in Vλµ . We call this
the Seiberg–Witten Floer homology of (Y, c).
Note that one can think of this finite dimensional flow as a perturbation of
the Seiberg–Witten flow on V. In [20], Marcolli and Wang used more standard
perturbations to define equivariant Seiberg–Witten Floer homology of rational
homology 3–spheres. A similar construction for all 3–manifolds is the object
of forthcoming work of Kronheimer and Mrowka [18]. It might be possible to
prove that our definition is equivalent to these by using a homotopy argument
as −λ, µ → ∞. However, such an argument would have to deal with both types
of perturbations at the same time. In particular, it would have to involve the
whole technical machinery of [20] or [18] in order to achieve a version of Morse–
Smale transversality, and this is not the goal of the present paper. We prefer
to work with SWF as it is defined here.
Geometry & Topology, Volume 7 (2003)
892
Ciprian Manolescu
In section 9 we construct a relative Seiberg–Witten invariant of four-manifolds
with boundary. Suppose that the boundary Y of a compact, oriented fourmanifold X is a (possibly empty) disjoint union of rational homology 3–spheres,
and that X has a spinc structure ĉ which restricts to c on Y. For any version
of Floer homology, one expects that the solutions of the Seiberg–Witten (or
instanton) equations on X induce by restriction to the boundary an element in
the Floer homology of Y. In our case, let Ind be the virtual index bundle over
the Picard torus H 1 (X; R)/H 1 (X; Z) corresponding to the Dirac operators on
X. If we write Ind as the difference between a vector bundle E with Thom space
T (E) and a trivial bundle Rα ⊕ Cβ , then let us denote T (Ind) = (T (E), α, β +
n(Y, c, g)) ∈ Ob C. The correction term n(Y, c, g) is included to make T (Ind)
metric independent. We will prove the following:
Theorem 2 Finite dimensional approximation of the Seiberg–Witten equations on X gives an equivariant stable homotopy class of maps:
Ψ(X, ĉ) ∈ {(T (Ind), b+
2 (X), 0), SWF(Y, c)}S 1 .
The invariant Ψ depends only on X and ĉ, up to canonical isomorphism.
In particular, when X is closed we recover the Bauer–Furuta invariant Ψ from
[4]. Also, in the general case by composing Ψ with the Hurewicz map we obtain
a relative invariant of X with values in the Seiberg–Witten–Floer homology of
Y.
When X is a cobordism between two 3–manifolds Y1 and Y2 with b1 = 0, we
will see that the invariant Ψ can be interpreted as a morphism DX between
SWF(Y1 ) and SWF(Y2 ), with a possible shift in degree. (We omit the spinc
structures from notation for simplicity.)
We expect the following gluing result to be true:
Conjecture 1 If X1 is a cobordism between Y1 and Y2 and X2 is a cobordism
between Y2 and Y3 , then
DX ∪X ∼
= DX ◦ DX .
1
2
2
1
A particular case of this conjecture (for connected sums of closed four-manifolds)
was proved in [5]. Note that if Conjecture 1 were true, this would give a construction of a “spectrum-valued topological quantum field theory” in 3+1 dimensions, at least for manifolds with boundary rational homology 3–spheres.
In section 10 we present an application of Theorem 2. We specialize to the case
of four-manifolds with boundary that have negative definite intersection form.
Geometry & Topology, Volume 7 (2003)
Seiberg–Witten–Floer stable homotopy types
893
For every integer r ≥ 0 we construct an element γr ∈ H 2r+1 (swfirr
>0 (Y, c, g, ν); Z),
where swfirr
is
a
metric
dependent
invariant
to
be
defined
in
section
8 (roughly,
>0
it equals half of the irreducible part of SWF.) We show the following bound,
which parallels the one obtained by Frøyshov in [12]:
Theorem 3 Let X be a smooth, compact, oriented 4–manifold such that
b+
2 (X) = 0 and ∂X = Y has b1 (Y ) = 0. Then every characteristic element
c ∈ H2 (X, ∂X)/Torsion satisfies:
b2 (X) + c2
≤ max inf −n(Y, c, g) + min{r|γr = 0} .
c
g,ν
8
Acknowledgements This paper was part of the author’s senior thesis. I am
extremely grateful to my advisor, Peter Kronheimer, for entrusting this project
to me, and for all his invaluable support. I would also like to thank Lars Hesselholt, Michael Hopkins, and Michael Mandell for having taken the time to
answer my questions on equivariant stable homotopy theory. I am grateful to
Tom Mrowka, Octavian Cornea, Dragos Oprea and Jake Rasmussen for helpful
conversations, and to Ming Xu for pointing out a mistake in a previous version. Finally, I would like to thank the Harvard College Research Program for
partially funding this work.
2
Seiberg–Witten trajectories
We start by reviewing a few basic facts about the Seiberg–Witten equations on
three-manifolds and cylinders. Part of our exposition is inspired from [16], [17],
and [18].
Let Y be an oriented 3–manifold endowed with a metric g and a spinc structure
c with spinor bundle W0 . Our orientation convention for the Clifford multiplication ρ : T Y → End (W0 ) is that ρ(e1 )ρ(e2 )ρ(e3 ) = 1 for an oriented frame ei .
Let L = det(W0 ), and assume that b1 (Y ) = 0. The fact that b1 (Y ) = 0 implies
the existence of a flat spinc connection A0 . This allows us to identify the affine
space of spinc connections A on W0 with iΩ1 (Y ) by the correspondence which
sends a ∈ iΩ1 (Y ) to A0 + a.
Let us denote by 6 ∂ a = ρ(a) + 6 ∂ : Γ(W0 ) → Γ(W0 ) the Dirac operator associated to the connection A0 + a. In particular, 6 ∂ = 6 ∂ 0 corresponds to the flat
connection.
Geometry & Topology, Volume 7 (2003)
894
Ciprian Manolescu
The gauge group G = Map(Y, S 1 ) acts on the space iΩ1 (Y ) ⊕ Γ(W0 ) by
u(a, φ) = (a − u−1 du, uφ).
It is convenient to work with the completions of iΩ1 (Y ) ⊕ Γ(W0 ) and G in the
L2k+1 and L2k+2 norms, respectively, where k ≥ 4 is a fixed integer. In general,
we denote the L2k completion of a space E by L2k (E).
The Chern–Simons–Dirac functional is defined on L2k+1 (iΩ1 (Y ) ⊕ Γ(W0 )) by
Z
Z
1
CSD(a, φ) = −
a ∧ da + hφ, 6 ∂ a φidvol .
2
Y
Y
R −1
1
We have CSD(u·(a, φ))−CSD(a, φ) = 2 Y u du∧da = 0 because H 1 (Y ; Z) =
0, so the CSD functional is gauge invariant. A simple computation shows that
its gradient (for the L2 metric) is the vector field
∇CSD(a, φ) = (∗da + τ (φ, φ), 6 ∂ a φ),
where τ is the bilinear form defined by τ (φ, ψ) = ρ−1 (φψ ∗ )0 and the subscript
0 denotes the trace-free part.
The Seiberg–Witten equations on Y are given by
∗da + τ (φ, φ) = 0, 6 ∂ a φ = 0,
so their solutions are the critical points of the Chern–Simons–Dirac functional.
A solution is called reducible if θ = 0 and irreducible otherwise.
The following result is well-known (see [17] for the analogue in four dimensions,
or see [16]):
Lemma 1 Let (a, φ) be a C 2 solution to the Seiberg–Witten equations on
Y. Then there exists a gauge transformation u such that u(a, φ) is smooth.
Moreover, there are upper bounds on all the C m norms of u(a, φ) which depend
only on the metric on Y.
Let us look at trajectories of the downward gradient flow of the CSD functional:
∂
x(t) = −∇CSD(x(t)).
(1)
∂t
Seiberg–Witten trajectories x(t) as above can be interpreted in a standard way
as solutions of the four-dimensional monopole equations on the cylinder R × Y.
A spinc structure on Y induces one on R × Y with spinor bundles W ± , and
a path of spinors φ(t) on Y can be viewed as a positive spinor Φ ∈ Γ(W + ).
Similarly, a path of connections A0 + a(t) on Y produces a spinc connection
x = (a, φ) : R → L2k+1 (iΩ1 (Y ) ⊕ Γ(W0 )),
Geometry & Topology, Volume 7 (2003)
Seiberg–Witten–Floer stable homotopy types
895
A on R × Y by adding a d/dt component to the covariant derivative. There
+
is a corresponding Dirac operator DA
: Γ(W + ) → Γ(W − ). Let us denote by
+
FA the self-dual part of half the curvature of the connection induced by A on
det(W ± ) and let us extend Clifford multiplication ρ̂ to 2–forms in the usual
way. Set σ(Φ, Φ) = ρ̂−1 (ΦΦ∗ )0 . The fact that x(t) = (a(t), φ(t)) satisfies (1)
can be written as
+
DA
Φ = 0, FA+ = σ(Φ, Φ).
These are exactly the four-dimensional Seiberg–Witten equations.
Definition 1 A Seiberg–Witten trajectory x(t) is said to be of finite type if
both CSD(x(t)) and kφ(t)kC 0 are bounded functions of t.
Before proving a compactness result for trajectories of finite type analogous to
Lemma 1, we need to define a useful concept. If (A, Φ) are a spinc connection
and a positive spinor on a comapct 4–manifold X, we say that the energy of
(A, Φ) is the quantity:
Z 1
1
s
E(A, Φ) =
|FA |2 + |∇A Φ|2 + |Φ|4 + |Φ|2 ,
2 X
4
4
where s denotes the scalar curvature. It is easy to see that E is gauge invariant.
In the case when X = [α, β] × Y and (A, Φ) is a Seiberg–Witten trajectory
x(t) = (a(t), φ(t)), t ∈ [α, β], the energy can be written as the change in the
CSD functional. Indeed,
Z β
CSD(x(α)) − CSD(x(β)) =
k(∂/∂t)a(t)k2L2 + k(∂/∂t)φ(t)k2L2 dt (2)
Zα
=
|∗da + τ (φ, φ)|2 + |6 ∂ a φ|2
ZX
1
s
=
|∗da|2 + |∇a φ|2 + |φ|4 + |φ|2 .
4
4
X
It is now easy to see that the last expression equals E(A, Φ). In the last step
of the derivation we have used the Weitzenböck formula.
We have the following important result for finite type trajectories:
Proposition 1 There exist Cm > 0 such that for any (a, φ) ∈ L2k+1 (iΩ1 (Y )⊕
Γ(W0 )) which is equal to x(t0 ) for some t0 ∈ R and some Seiberg–Witten
trajectory of finite type x : R → L2k+1 (iΩ1 (Y ) ⊕ Γ(W0 )), there exists (a0 , φ0 )
smooth and gauge equivalent to (a, φ) such that k(a0 , φ0 )kC m ≤ Cm for all
m > 0.
Geometry & Topology, Volume 7 (2003)
896
Ciprian Manolescu
First we must prove:
Lemma 2 Let X be a four-dimensional Riemannian manifold with boundary
such that H 1 (X; R) = 0. Denote by ν the unit normal vector to ∂X. Then there
is a constant K > 0 such that for any A ∈ Ω1 (X) continuously differentiable,
with A(ν) = 0 on ∂X, we have:
Z
Z
2
|A| < K
(|dA|2 + |d∗ A|2 ).
X
X
Proof Assume there is no such K . Then we can find a sequence of normalized
An ∈ Ω1 with
Z
Z
2
|An | = 1, (|dAn |2 + |d∗ An |2 ) → 0.
X
X
The additional condition An (ν) = 0 allows us to integrate by parts in the
Weitzenböck formula to obtain:
Z
Z
2
|∇An | + hRic(An ), An i =
|dAn |2 + |d∗ An |2 .
X
X
Since Ric is a bounded tensor of An we obtain a uniform bound on k∇An kL2 .
By replacing An with a subsequence we can assume that An converge weakly
in L21 norm to some A such that dA = d∗ A = 0. Furthermore, since the
restriction map from L21 (X) to L2 (∂X) is compact, we can also assume that
An |∂X → A|∂X in L2 (∂X). Hence A(ν) = 0 on ∂X (Neumann boundary value
condition) and A is harmonic on X, so A = 0. This contradicts the strong L2
convergence An → A and the fact that kAn kL2 = 1.
Proof of Proposition 1 We start by deriving a pointwise bound on the spinorial part. Consider a trajectory of finite type x = (a, φ) : R → L2k+1 (iΩ1 (Y )⊕
Γ(W0 )). Let S be the supremum of the pointwise norm of φ(t) over R × Y.
If |φ(t)(y)| = S for some (y, t) ∈ R × Y, since φ(t) ∈ L25 ⊂ C 2 , we have
∆|φ|2 ≥ 0 at that point. Here ∆ is the four-dimensional Laplacian on R × Y.
By the standard compactness argument for the Seiberg–Witten equations [17],
we obtain an upper bound for |φ| which depends only on the metric on Y.
If the supremum is not attained, we can find a sequence (yn , tn ) ∈ R × Y with
|φ(tn )(yn )| → S. Without loss of generality, by passing to a subsequence we
can assume that yn → y ∈ Y and tn+1 > tn + 2 (hence tn → ∞). Via a
reparametrization, the restriction of x to each interval [tn − 1, tn + 1] can be
interpreted as a solution (An , Φn ) of the Seiberg–Witten equations on X =
[−1, 1] × Y. The finite type hypothesis and formula (2) give uniform bounds on
Geometry & Topology, Volume 7 (2003)
Seiberg–Witten–Floer stable homotopy types
897
|Φn | and kdAn kL2 . Here we identify connections with 1–forms by comparing
them to the standard product flat connection.
We can modify (An , Φn ) by a gauge transformation on X so that we obtain
d∗ An = 0 on X and An (∂/∂t) = 0 on ∂X. Using Lemma 2 we get a uniform
bound on kAn kL2 . After this point the Seiberg–Witten equations
+
DA
Φ = 0, d+ An = σ(Φn , Φn )
n n
provide bounds on all the Sobolev norms of An |X 0 and Φn |X 0 by elliptic bootstrapping. Here X 0 could be any compact subset in the interior of X, for
example [−1/2, 1/2] × Y.
Thus, after to passing to a subsequence we can assume that (An , Φn )|X 0 converges in C ∞ to some (A, Φ), up to some gauge transformations. Note that
the energies on X 0
Z tn +1/2
1
1 E 0 (An , Φn ) = CSD(x(tn − ))−CSD(x(tn + )) =
k(∂/∂t)x(t)k2L2 dt
2
2
tn −1/2
P 0
are positive, while the series
n E (An , Φn ) is convergent because CSD is
0
bounded. It follows that E (An , Φn ) → 0 as n → ∞, so E 0 (A, Φ) = 0. In
temporal gauge on X 0 , (A, Φ) must be of the form (a(t), φ(t)), where a(t) and
φ(t) are constant in t, giving a solution of the Seiberg–Witten equations on Y.
By Lemma 1, there is an upper bound for |φ(0)(y)| which depends only on Y.
Now φ(tn )(yn ) converges to φ(0)(y) up to some gauge transformation, hence
the upper bound also applies to limn |φ(tn )(yn )| = S.
Therefore, in all cases we have a uniform bound kφ(t)kC 0 ≤ C for all t and for
all trajectories.
The next step is to deduce a similar bound for the absolute value of CSD(x(t)).
Observe that CSD(x(t)) > CSD(x(n)) for all n sufficiently large. As before,
we interpret the restriction of x to each interval [n − 1, n + 1] as a solution
of the Seiberg–Witten equations on [−1, 1] × Y. Then we find that a subsequence of these solutions restricted to X 0 converges to some (A, Φ) in C ∞ .
Also, (A, Φ) must be constant in temporal gauge. We deduce that a subsequence of CSD(x(n)) converges to CSD(a, φ), where (a, φ) is a solution of
the Seiberg–Witten equations on Y. Using Lemma 1, we get a lower bound for
CSD(x(t)). An upper bound can be obtained similarly.
Now let us concentrate on a specific x(t0 ). By a linear reparametrization, we
can assume t0 = 0. Let X = [−1, 1] × Y. Then (A, Φ) = (a(t), φ(t)) satisfies the
4–dimensional Seiberg–Witten equations. The formula (2) and the bounds on
|φ| and |CSD| imply a uniform bound on kdAkL2 . Via a gauge transformation
Geometry & Topology, Volume 7 (2003)
898
Ciprian Manolescu
on X we can assume that d∗ A = 0 on X and An (ν) = 0 on ∂X. By Lemma 2
we obtain a bound on kAkL2 and then, by elliptic bootstrapping, on all Sobolev
norms of A and Φ. The desired C m bounds follow.
The same proof works in the setting of a half-trajectory of finite type glued to
a four-manifold with boundary. We state here the relevant result, which will
prove useful to us in section 9.
Proposition 2 Let X be a Riemannian four-manifold with a cylindrical end
U isometric to (0, ∞) × R, and such that X \ U is compact. Let t > 0 and
Xt = X \ ([t, ∞) × R). Then there exist Cm,t > 0 such that any monopole on
X which is gauge equivalent to a half-trajectory of finite type over U is in fact
gauge equivalent over Xt to a smooth monopole (A, Φ) such that k(A, Φ)kC m ≤
Cm,t for all m > 0.
3
Projection to the Coulomb gauge slice
Let G0 beRthe group of “normalized” gauge transformations, ie, u : Y → S 1 , u =
eiξ with Yj ξ = 0 for any connected component Yj of Y. It will be helpful to
work on the space
V = i ker d∗ ⊕ Γ(W0 ).
For (a, φ) ∈ iΩ1 (Y ) ⊕ Γ(W0 ), there is a unique element of V which is equivalent
to (a, φ) by a transformation in G0 . We call this element the Coulomb projection
of (a, φ).
Denote by π the orthogonal projection from Ω1 (Y ) to ker d∗ . The space V inherits a metric g̃ from the L2 inner product on iΩ1 (Y ) ⊕ Γ(W0 ) in the following
way: given (b, ψ) a tangent vector at (a, φ) ∈ V, we set
k(b, ψ)kg̃ = k(b, ψ) + (−idξ, iξφ)kL2
where ξ ∈ G0 is such that (b − idξ, ψ + iξφ) is in Coulomb gauge, ie,
d∗ (b − idξ) + 2iRehiφ, ψ + iξφi = 0.
The trajectories of the CSD functional restricted to V in this metric are the
same as the Coulomb projections of the trajectories of the CSD functional on
iΩ1 (Y ) ⊕ Γ(W0 ).
∗ ⊥
For φ ∈ Γ(W0 ), note that (1 − π) ◦ τ (φ, φ) ∈ (ker
R d ) = Im d. Define ξ(φ) :
Y → R by dξ(φ) = i(1 − π) ◦ τ (φ, φ) and Yj ξ(φ) = 0 for all connected
components Yj ⊂ Y.
Geometry & Topology, Volume 7 (2003)
899
Seiberg–Witten–Floer stable homotopy types
Then the gradient of CSD|V in the g̃ metric can be written as l + c, where
l, c : V → V are given by
l(a, φ) = (∗da, 6 ∂φ)
c(a, φ) = π ◦ τ (φ, φ), ρ(a)φ − iξ(φ)φ .
Thus from now on we can concentrate on trajectories x : R → V, (∂/∂t)x(t) =
−(l + c)x(t). More generally, we can look at such trajectories with values in the
L2k+1 completion of V. Note that l + c : L2k+1 (V ) → L2k (V ) is a continuous map.
We construct all Sobolev norms on V using l as the differentiation operator:
m Z
X
2
kvkL2m (V ) =
|lj (v)|2 dvol.
j=0
Y
Consider such trajectories x : R → L2k+1 (V ), k ≥ 4. Assuming they are of
finite type, from Proposition 1 we know that they are locally the projections of
smooth trajectories living in the ball of of radius Cm in the C m norm, for each
m. We deduce that x(t) ∈ V for all t, x is smooth in t and there is a uniform
bound on kx(t)kC m for each m.
4
Finite dimensional approximation
In this section we use Furuta’s technique to prove an essential compactness
result for approximate Seiberg–Witten trajectories.
Note that the operator l defined in the previous section is self-adjoint, so has
only real eigenvalues. In the standard L2 metric, let p̃µλ be the orthogonal
projection from V to the finite dimensional subspace Vλµ spanned by the eigenvectors of l with eigenvalues in the interval (λ, µ].
It is useful to consider a modification of the projections so that we have a
continuous family of maps, as in [14]. Thus let β : R → [0, ∞) be a smooth
function so that β(x) > 0 ⇐⇒ x ∈ (0, 1) and the integral of β is 1. For each
−λ, µ > 1, set
Z
pµλ =
1
0
µ−τ
β(τ )p̃λ+τ
dτ.
Now pµλ : V → V varies continuously in λ and µ. Also Vλµ = Im(pµλ ), except
when µ is an eigenvalue. Let us modify the definition of Vλµ slightly so that
it is always the image of pµλ ). (However, we only do that for µ > 1; later on,
when we talk about Vλλ0 for λ0 < λ ≤ 0, for technical reasons we still want it to
be the span of eigenspaces with eigenvalues in (λ0 , λ].)
Geometry & Topology, Volume 7 (2003)
900
Ciprian Manolescu
Let k ≥ 4. Then c : L2k+1 (V ) → L2k (V ) is a compact map. This follows
from the following facts: ξ maps L2k+1 to L2k+1 ; the Sobolev multiplication
L2k+1 × L2k+1 → L2k+1 is continuous; and the inclusion L2k+1 → L2k is compact.
A useful consequence of the compactness of c is that we have
k(1 − pµλ )c(x)kL2 → 0
k
when −λ, µ → ∞, uniformly in x when x is bounded in L2k+1 (V ).
Let us now denote by B(R) the open ball of radius R in L2k+1 (V ). We know
that there exists R > 0 such that all the finite type trajectories of l + c are
inside B(R).
Proposition 3 For any −λ and µ sufficiently large, if a trajectory x : R →
L2k+1 (Vλµ ),
∂
(l + pµλ c)(x(t)) = − x(t)
∂t
satisfies x(t) ∈ B(2R) for all t, then in fact x(t) ∈ B(R) for all t.
We organize the proof in three steps.
Step 1 Assume that the conclusion is false, so there exist sequences −λn , µn →
∞ and corresponding trajectories xn : R → B(2R) satisfying
∂
xn (t),
∂t
and (after a linear reparametrization) xn (0) 6∈ B(R). Let us denote for simplicity πn = pµλnn and π n = 1 − πn . Since l and c are bounded maps from L2k+1 (V )
to L2k (V ), there is a uniform bound
(l + pµλnn c)(xn (t)) = −
k
∂
xn (t)kL2
k
∂t
≤ kl(xn (t))kL2 + kπn c(xn (t))kL2
k
k
≤ kl(xn (t))kL2 + kc(xn (t))kL2
k
≤ Ckxn (t)kL2
k+1
k
≤ 2CR
for some constant C, independent of n and t. Therefore xn are equicontinuous
in L2k norm. They also sit inside a compact subset B 0 of L2k (V ), the closure
of B(2R) in this norm. After extracting a subsequence we can assume by the
Arzela–Ascoli theorem that xn converge to some x : R → B 0 , uniformly in L2k
norm over compact sets of t ∈ R. Letting n go to infinity we obtain
−
∂
xn (t) = (l + c)xn (t) − π n c(xn (t)) → (l + c)x(t)
∂t
Geometry & Topology, Volume 7 (2003)
Seiberg–Witten–Floer stable homotopy types
901
in L2k−1 (V ), uniformly on compact sets of t. From here we get that
Z t
Z t
∂
xn (t) =
xn (s)ds → − (l + c)(x(s))ds.
0 ∂t
0
On the other hand, we also know that xn (t) converges to x(t), so
∂
x(t).
∂t
The Chern–Simons–Dirac functional and the pointwise norm of the spinorial
part are bounded on the compact set B 0 . We conclude that x(t) is the Coulomb
projection of a finite type trajectory for the usual Seiberg–Witten equations on
Y. In particular, x(t) is smooth, both on Y and in the t direction. Also
x(0) ∈ B(R). Thus
kx(0)kL2 < R.
(3)
(l + c)x(t) = −
k+1
We seek to obtain a contradiction between (3) and the fact that xn (0) 6∈ B(R)
for any n.
Step 2 Let W be the vector space of trajectories x : [−1, 1] → V, x(t) =
(a(t), φ(t)). We can introduce Sobolev norms L2m on this space by looking at
a(t), φ(t) as sections of bundles over [−1, 1] × Y.
We will prove that xn (t) → x(t) in L2k (W ).
To do this, it suffices to prove that for every j(0 ≤ j ≤ k) we have
∂ j
∂ j
xn (t) →
x(t)
∂t
∂t
in L2k−j (V ), uniformly in t, for t ∈ [−1, 1]. We already know this statement to
be true for j = 0, so we proceed by induction on j.
Assume that
∂ s
xn (t) →
∂ s
x(t) in L2k−s (V ),
∂t
∂t
uniformly in t, for all s ≤ j. Then
∂ j+1
∂ j
−
(l + πn c)(xn (t)) − (l + c)(x(t))
(xn (t) − x(t)) =
∂t
∂t
∂ j
∂ j
∂ j
=l
(xn (t)−x(t)) +πn
(c(xn (t))−c(x(t))) −π n
c(x(t)) .
∂t
∂t
∂t
Here we have used the linearity of l, πn and π n . We discuss each of the three
terms in the sum above separately and prove that each of them converges to 0
Geometry & Topology, Volume 7 (2003)
902
Ciprian Manolescu
in L2k−j−1 uniformly in t. For the first term this is clear, because l is a bounded
linear map from L2k−j to L2k−j−1 .
For the third term, y(t) = (∂/∂t)j c(x(t)) is smooth over [−1, 1] × Y by what
we showed in Step 1, and kπ n y(t)k → 0 for each t ∈ [−1, 1] by the spectral
theorem. Here the norm can be any Sobolev norm, in particular L2k−j−1 . The
convergence is uniform in t because of smoothness in the t direction. Indeed,
assume that we can find > 0 and tn ∈ [−1, 1] so that kπ n y(tn )k ≥ for all
n. By going to a subsequence we can assume tn → t ∈ [−1, 1]. Then
≤ kπ n y(tn )k ≤ kπ n (y(tn ) − y(t))k + kπ n y(t)k
This is a contradiction. The last expression converges to zero because the first
term is less or equal to ky(tn ) − y(t)k.
All that remains is to deal with the second term. Since kπn yk ≤ kyk for every
Sobolev norm, it suffices to show that
∂ j
∂ j
c(xn (t)) →
c(x(t)) in L2k−j−1 (V )
∂t
∂t
uniformly in t. In fact we will prove a stronger L2k−j convergence. Note that
c(xn ) is quadratic in xn = (an , φn ) except for the term −iξ(φn )φn . Expanding
(∂/∂t)j c(xn (t)) by the Leibniz rule, we get expressions of the form
∂ s
∂ j−s
zn (t) ·
wn (t),
∂t
∂t
where zn , wn are either ξ(φn ) or local coordinates of xn . Assume they are both
local coordinates of xn . By the inductive hypothesis, we have (∂/∂t)s zn (t) →
(∂/∂t)s z(t) in L2k−s and (∂/∂t)j−s wn (t) → (∂/∂t)j−s w(t) in L2k−j+s, both
uniformly in t. Note that max (k − s, k − j + s) ≥ (k − s + k − j + s)/2 =
k − (j/2) ≥ k/2 ≥ 2. Therefore there is a Sobolev multiplication
L2k−s × L2k−j+s → L2min(k−s,k−j+s)
and the last space is contained in L2k−j . It follows that
∂ s
∂t
zn (t) ·
∂ j−s
∂t
wn (t) →
∂ s
∂t
z(t) ·
∂ j−s
∂t
w(t)
in L2k−j , uniformly in t.
The same is true when one or both of zn , wn are ξ(φn ). Clearly it is enough to
show that (∂/∂t)s ξ(φn (t)) → (∂/∂t)s ξ(φ(t)) in L2k−s , uniformly in t, for s ≤ j.
But from the discussion above we know that this is true if instead of ξ we had
Geometry & Topology, Volume 7 (2003)
Seiberg–Witten–Floer stable homotopy types
903
dξ, because this is quadratic in φn . Hence the convergence is also true for ξ.
This concludes the inductive step.
Step 3 The argument in this part is based on elliptic bootstrapping for the
equations on X = [−1, 1] × Y. Namely, the operator D = −∂/∂t − l acting on
W is Fredholm (being the restriction of an elliptic operator). We know that
Dxn (t) = πn c(xn (t)),
where xn (t) → x(t) in L2k (W ). We prove by induction on m ≥ k that xn (t) →
x(t) in L2m (Wm ), where Wm is the restriction of W to Xm = Im × Y and
Im = [−1/2 − 1/m, 1/2 + 1/m]. Assume this is true for m and we prove it for
m + 1. The elliptic estimate gives
kxn (t)−x(t)kL2m+1 (Wm+1 ) ≤ C kD(xn (t)−x(t))kL2m (Wm ) +kxn (t)−x(t)kL2m (Wm )
≤ C kπn c(xn (t)) − πn c(x(t))k + kπ n c(x(t))k + kxn (t) − x(t)k .
(4)
In the last expression all norms are taken in the L2m (Wm ) norm. We prove that
each of the three terms converges to zero when n → ∞. This is clear for the
third term from the inductive hypothesis.
For the first term, note that πn c is quadratic in xn (t), apart from the term
involving ξ(φn (t)). Looking at xn (t) as L2m sections of a bundle over Xm , the
Sobolev multiplication L2m ×L2m → L2m tells us that the quadratic terms are continuous maps from L2m (Wm ) to itself. From here we also deduce that dξ(φn (t)),
which is quadratic in its argument, converges to dξ(φ(t)). By integrating over
Im we get:
Z
Z
kξ(φn (t)) − ξ(φ(t))kL2m+1 (V ) ≤
C · kdξ(φn (t)) − dξ(φ(t))kL2m (V )
Im
Im
The right hand side of this inequality converges to zero as n → ∞, hence so does
the left hand side. Furthermore, the same is true if we replace ξ by (∂/∂t)s ξ
and m by m−s. Therefore, ξn (φn (t)) → ξ(φ(t)) in L2m (Wm ), so by the Sobolev
multiplication the first term in (4) converges to zero.
Finally, for the second term in (4), recall from Step 1 that c(x(t)) is smooth.
Hence π n (∂/∂t)s c(x(t)) converges to zero in L2m (V ), for each t and for all s ≥ 0.
The convergence is uniform in t because of smoothness in the t direction, by
an argument similar to the one in Step 2. We deduce that
kπ n (∂/∂t)s c(x(t))kL2m (Wm ) → 0
as well.
Geometry & Topology, Volume 7 (2003)
904
Ciprian Manolescu
Now we can conclude that the inductive step works, so xn (t) → x(t) in L2m (W 0 )
for all m if we take W 0 to be the restriction of W to [−1/2, 1/2] × Y. Convergence in all Sobolev norms means C ∞ convergence, so in particular xn (0) ∈
C ∞ (V ) and xn (0) → x(0) in C ∞ . Hence
kxn (0)kL2
k+1 (V
)
→ kx(0)kL2
k+1 (V
).
We obtain a contradiction with the fact that xn (0) 6∈ B(R), so kxn (0)kL2 (V ) ≥
k+1
R, while kx(0)kL2 (V ) < R.
k+1
5
The Conley index
The Conley index is a well-known invariant in dynamics, developed by C. Conley in the 70’s. Here we summarize its construction and basic properties, as
presented in [7] and [24].
Let M be a finite dimensional manifold and ϕ a flow on M, ie, a continuous
map ϕ : M × R → M, (x, t) → ϕt (x), satisfying ϕ0 = id and ϕs ◦ ϕt = ϕs+t .
For a subset A ⊂ M we define
A+ = {x ∈ A : ∀t > 0, ϕt (x) ∈ A};
A− = {x ∈ A : ∀t < 0, ϕt (x) ∈ A};
Inv A = A+ ∩ A− .
It is easy to see that all of these are compact subsets of A, provided that A
itself is compact.
A compact subset S ⊂ M is called an isolated invariant set if there exists a
compact neighborhood A such that S = Inv A ⊂ int(A). Such an A is called
an isolating neighborhood of S. It follows from here that Inv S = S.
A pair (N, L) of compact subsets L ⊂ N ⊂ M is said to be an index pair for
S if the following conditions are satisfied:
(1) Inv (N \ L) = S ⊂ int(N \ L);
(2) L is an exit set for N, ie, for any x ∈ N and t > 0 such that ϕt (x) 6∈ N,
there exists τ ∈ [0, t) with ϕτ (x) ∈ L.
(3) L is positively invariant in N, ie, if for x ∈ L and t > 0 we have
ϕ[0,t] (x) ⊂ N, then in fact ϕ[0,t] (x) ⊂ L.
Geometry & Topology, Volume 7 (2003)
Seiberg–Witten–Floer stable homotopy types
905
Consider an isolated invariant set S ⊂ M with an isolating neighborhood A.
The fundamental result in Conley index theory is that there exists an index pair
(N, L) for S such that N ⊂ A. We prove this theorem in a slightly stronger
form which will be useful to us in section 9; the proof is relegated to Appendix
A:
Theorem 4 Let S ⊂ M be an isolated invariant set with a compact isolating neighborhood A, and let K1 , K2 ⊂ A be compact sets which satisfy the
following conditions:
(i) If x ∈ K1 ∩ A+ , then ϕt (x) 6∈ ∂A for any t ≥ 0;
(ii) K2 ∩ A+ = ∅.
Then there exists an index pair (N, L) for S such that K1 ⊂ N ⊂ A and
K2 ⊂ L.
Given an isolated invariant set S with index pair (N, L), one defines the Conley
index of S to be the pointed homotopy type
I(ϕ, S) = (N/L, [L]).
The Conley index has the following properties:
(1) It depends only on S. In fact, there are natural pointed homotopy equivalences between the spaces N/L for different choices of the index pair.
(2) If ϕi is a flow on Mi , i = 1, 2, then I(ϕ1 × ϕ2 , S1 × S2 ) ∼
= I(ϕ1 , S1 ) ∧
I(ϕ2 , S2 ).
(3) If A is an isolating neighborhood for St = Inv A for a continuous family of
flows ϕt , t ∈ [0, 1], then I(ϕ0 , S0 ) ∼
= I(ϕ1 , S1 ). Again, there are canonical
homotopy equivalences between the respective spaces.
By abuse of notation, we will often use I to denote the pointed space N/L,
and say that N/L “is” the Conley index.
To give a few examples of Conley indices, for any flow I(ϕ, ∅) is the homotopy
type of a point. If p is a nondegenerate critical point of a gradient flow ϕ on M,
then I(ϕ, {p}) ∼
= S k , where k is the Morse index of p. More generally, when ϕ
is a gradient flow and S is an isolated invariant set composed of critical points
and trajectories between them satisfying the Morse-Smale condition, then one
can compute a Morse homology in the usual way (as in [25]), and it turns out
that it equals H̃∗ (I(ϕ, S)).
Geometry & Topology, Volume 7 (2003)
906
Ciprian Manolescu
Another useful property of the Conley index is its behavior in the presence of
attractor-repeller pairs. Given a subset A ⊂ M, we define its α–limit set and
its ω –limit set as:
\
\
α(A) =
ϕ(−∞,t] (A); ω(A) =
ϕ[t,∞) (A).
t<0
t>0
If S is an isolated invariant set, a subset T ⊂ S is called a repeller (resp.
attractor ) relative to S is there exists a neighborhood U of T in S such that
α(U ) = T (resp. ω(U ) = T ). If T ⊂ S is an attractor, then the set T ∗ of
x ∈ S such that ω(x) ∩ T = ∅ is a repeller in S, and (T, T ∗ ) is called an
attractor-repeller pair in S. To give an example, let S be a set of critical points
and the trajectories between them in a gradient flow ϕ generated by a Morse
function f on M. Then, for some a ∈ R, we could let T ⊂ S be the set of
critical points x for which f (x) ≤ a, together with the trajectories connecting
them. In this case T ∗ is the set of critical points x ∈ S for which f (x) > a,
together with the trajectories between them.
In general, for an attractor-repeller pair (T, T ∗ ) in S, we have the following:
Proposition 4 Let A be an isolating neighborhood for S. Then there exist
compact sets N3 ⊂ N2 ⊂ N1 ⊂ A such that (N1 , N2 ), (N1 , N3 ), and (N2 , N3 )
are index pairs for T ∗ , S, and T, respectively. Hence there is a coexact sequence:
I(ϕ, T ) → I(ϕ, S) → I(ϕ, T ∗ ) → ΣI(ϕ, T ) → ΣI(ϕ, S) → . . .
Finally, we must note that an equivariant version of the Conley index was
constructed by A. Floer in [11] and extended by A. M. Pruszko in [23]. Let G
be a compact Lie group; in this paper we will be concerned only with G = S 1 .
If the flow ϕ preserves a G–symmetry on M and S is an isolated invariant set
which is also invariant under the action of G, then one can generalize Theorem 4
to prove the existence of an G–invariant index pair with the required properties.
The resulting Conley index IG (ϕ, S) is an element of G–equivariant pointed
homotopy type. It has the same three basic properties described above, as well
as a similar behavior in the presence of attractor-repeller pairs.
6
Construction of the invariant
Let us start by defining the equivariant graded suspension category C. Our
construction is inspired from [1], [9], and [19]. However, for the sake of simplicity
we do not work with a universe, but we follow a more classical approach. There
Geometry & Topology, Volume 7 (2003)
907
Seiberg–Witten–Floer stable homotopy types
are several potential dangers in doing this in an equivariant setting (see [1],
[19]). However, in our case the Burnside ring A(S 1 ) = Z is particularly simple,
and it turns out that our construction does not involve additional complications
compared to its non-equivariant analogue in [27] and [21].
We are only interested in suspensions by the representations R and C of S 1 .
Thus, the objects of C are triplets (X, m, n), where X is a pointed topological
space with an S 1 –action, m ∈ Z and n ∈ Q. We require that X has the S 1 –
homotopy type of a S 1 –CW complex (this is always true for Conley indices
on manifolds). The set {(X, m, n), (X 0 , m0 , n0 )}S 1 of morphisms between two
objects is nonempty only for n − n0 ∈ Z and in this case it equals
0
0
{(X, m, n), (X 0 , m0 , n0 )}S 1 = colim [(Rk ⊕Cl )+ ∧X, (Rk+m−m ⊕Cl+n−n )+ ∧X 0 ]S 1
The colimit is taken over k, l ∈ Z. The maps that define the colimit are given
by suspensions, ie, smashing on the left at each step with either id(R+ ∧∗∧∗) or
id(∗∧C+ ∧∗) .
Inside of C we have a subcategory C0 consisting of the objects (X, 0, 0). We
usually denote such an object by X. Also, in general, if X 0 = (X, m, n) is any
object of C, we write (X 0 , m0 , n0 ) for (X, m + m0 , n + n0 ).
Given a finite dimensional vector space E with trivial S 1 action, we can define
the desuspension of X ∈ Ob C0 by E to be Σ−E X = (E + ∧ X, 2 dim E, 0).
Alternatively, we can set Σ−E X = ΩE X, the set of pointed maps from E +
to X. It is easy to check that these two definitions give the same object in
C, up to canonical isomorphism. Similarly, when E has free S 1 action apart
from the origin, one can define Σ−E X = ΩE X. This is naturally isomorphic to
(X, 0, dimC E), because E has a canonical orientation coming from its complex
structure.
Now recall the notations from section 4. We would like to consider the downward gradient flow of the Chern–Simons–Dirac functional on Vλµ in the metric
g̃. However, there could be trajectories that go to infinity in finite time, so this
is not well-defined. We need to take a compactly supported, smooth cut-off
function uµλ on Vλµ which is identically 1 on B(3R), where R is the constant
0
from Proposition 3. For consistency purposes we require uµλ = uµλ0 |V µ for λ0 ≤ λ
λ
and µ0 ≥ µ. Now for each λ and µ the vector field uµλ · (l + pµλ c) is compactly
supported, so it generates a well-defined flow ϕµλ on Vλµ .
From Proposition 3 we know that there exist −λ∗ , µ∗ > 1 such that for all
λ ≤ λ∗ , µ ≥ µ∗ , all trajectories of ϕµλ inside B(2R) are in fact contained in
B(R). It follows that Inv Vλµ ∩ B(2R) = Sλµ , the compact union of all such
trajectories, and Sλµ is an isolated invariant set.
Geometry & Topology, Volume 7 (2003)
908
Ciprian Manolescu
There is an S 1 symmetry in our case as a result of the division by G0 rather than
the full gauge group. We have the following S 1 action on V : eiθ ∈ S 1 sends
(a, φ) to (a, eiθ φ). The maps l and c are equivariant, and there is an induced
S 1 –action on the spaces Vλµ . Since both ϕµλ and Sλµ are invariant under the S 1
action, using the notion of equivariant Conley index from the previous section
we can set
Iλµ = IS 1 (ϕµλ , Sλµ ).
It is now the time to explain why we desuspended by Vλ0 in the definition of
SWF(Y, c) from the introduction. We also have to figure out what the value of
n(Y, c, g) should be.
One solution of the Seiberg–Witten equations on Y is the reducible θ = (0, 0).
Let X be a simply-connected oriented Riemannian four-manifold with boundary Y. Suppose that a neighborhood of the boundary is isometric to [0, 1] × Y
such that ∂X = {1} × Y. Choose a spinc structure on X which extends the one
on Y and let L̂ be its determinant line bundle. Let  be a smooth connection
on L̂ such that on the end it equals the pullback of the flat connection A0 on Y.
Then we can define c1 (L̂)2 ∈ Q in the following way. Let N be the cardinality
of H1 (Y ; Z). Then the exact sequence
∗
j
Hc2 (X) −−−−→ H 2 (X) −−−−→ H 2 (Y ) ∼
= H1 (Y )
tells us that N c1 (L̂) = j ∗ (α) for some α ∈ Hc2 (X). Using the intersection form
induced by Poincaré duality
Hc2 (X) × H 2 (X) → Z
we set
1
Z.
N
Denote by D± the Dirac operators on X coupled with the connection Â, with
Â
spectral boundary conditions as in [2]. One can look at solutions of the Seiberg–
Witten equations on X which restrict to θ on the boundary. The space M(X, θ)
of such solutions has a “virtual dimension”
χ(X) + σ(X) + 1
v.dim M(X, θ) = 2indC (D + ) −
.
(5)
Â
2
Here χ(X) and σ(X) are the Euler characteristic and the signature of X,
respectively.
c1 (L̂)2 = (α · c1 (L̂))/N ∈
In Seiberg–Witten theory, when one tries to define a version of Floer homology
it is customary to assign to the reducible a real index equal to
c1 (L̂)2 − (2χ(X) + 3σ(X) + 2)
− v.dim M(X, θ)
4
Geometry & Topology, Volume 7 (2003)
Seiberg–Witten–Floer stable homotopy types
909
On the other hand, if one were to compute the homology of the Conley index
Iλµ by a Morse homology recipe (counting moduli spaces of gradient flow lines),
the Morse index of (0, 0) would be different. In fact we can approximate l + pµλ c
near (0, 0) by its linear part l. The Morse index is then the number of negative
eigenvalues of l on Vλµ , which is the dimension of Vλ0 . (Our convention is to
also count the zero eigenvalues.) To account for this discrepancy in what the
index of θ should be, we need to desuspend by Vλ0 ⊕ Cn(Y,c,g) where it is natural
to set
1
c1 (L̂)2 − (2χ(X) + 3σ(X) + 2) n(Y, c, g) =
v.dim M(X, θ) −
.
2
4
We can simplify this expression using (5):
c1 (L̂)2 − σ(X)
.
8
1
We have n(Y, c, g) ∈ 8N
Z, where N is the cardinality of H1 (Y ; Z).
+
n(Y, c, g) = indC (DA
)−
(6)
Moreover, if Y is an integral homology sphere the intersection form on H2 (X)
is unimodular and this implies that c1 (L̂)2 ≡ σ(X) mod 8 (see for example
[15]). Therefore in this case n(Y, c, g) is an integer.
In general, we need to see that n(Y, c, g) does not depend on X. We follow [22]
and express it in terms of two eta invariants of Y. First, the index theorem of
Atiyah, Patodi and Singer for four-manifolds with boundary [2] gives
Z η − k(6 ∂)
1
1
dir
indC (D + ) =
− p1 + c1 (Â)2 +
.
(7)
Â
8 X
3
2
i
Here p1 and c1 (Â) = 2π
FÂ are the Pontryagin and Chern forms on X, while
k(6 ∂) = dim ker 6 ∂ = dim ker l as H1 (Y ; R) = 0. The eta invariant of a selfadjoint elliptic operator D on Y is defined to be the value at 0 of the analytic
continuation of the function
X
ηD (s) =
sign(λ)|λ|−s ,
λ6=0
where λ runs over the eigenvalues of D. In our case ηdir = η6∂ (0).
Let us also introduce the odd signature operator on Ω1 (Y ) ⊕ Ω0 (Y ) by
∗d −d
sign =
.
−d∗ 0
Then the signature theorem for manifolds with boundary [2] gives
Z
1
σ(X) =
p1 − ηsign .
3 X
Geometry & Topology, Volume 7 (2003)
(8)
910
Ciprian Manolescu
Putting (6), (7), and (8) together we obtain
Z
η − k(6 ∂) c (L̂)2 − σ(X)
1
1
1
dir
n(Y, c, g) =
− p1 + c1 (Â)2 +
−
8 X 3
2
8
1
ηsign =
ηdir − k(6 ∂) −
.
2
4
Warning Our sign conventions are somewhat different from those in [2]. In
our setting the manifold X has its boundary “on the right,” so that ∂/∂t on
[0, 1] × Y is the outward normal. Atiyah, Patodi, and Singer formulated their
theorem using the inward normal, so in order to be consistent we have applied
their theorem with −6 ∂ as the operator on the 3–manifold.
7
Proof of the main theorem
It is now the time to use the tools that we have developed so far to prove
Theorem 1 announced in the introduction. Recall that we are interested in
0
comparing the spectra (Σ−Vλ Iλµ , 0, n(Y, c, g)). We will denote by mλ the dimension of the real part of Vλ0 (coming from eigenspaces of ∗d), and by nλ the
complex dimension of the spinorial part of Vλ0 .
Proof of Theorem 1 First let us keep the metric on Y fixed and prove that
0
(Σ−Vλ Iλµ , 0, n(Y, c, g)) are naturally isomorphic for different λ ≤ λ∗ , µ ≥ µ∗ . In
0
fact we just need to do this for Σ−Vλ Iλµ , because n(Y, c, g) does not depend on
λ and µ.
It is not hard to see that for any λ and µ, the finite energy trajectories of
0
l + pµλcpµλ are contained in Vλµ . Let λ0 ≤ λ ≤ λ∗ , µ0 ≥ µ ≥ µ∗ . Then B(2R)∩ Vλµ0
is an isolating neighborhood for all Sab , a ∈ [λ0 , λ], b ∈ [µ, µ0 ]. By Property 3 of
0
0
the Conley index, for ϕ̃µλ the flow of uµλ0 · (l + pµλ c) on Vλµ0 ,
0
µ
µ
Iλµ0 ∼
= IS 1 (ϕ̃λ , Sλ ).
0
Let Vλµ0 = Vλµ ⊕ V̄ so that V̄ is the orthogonal complement of Vλµ in the L2
metric, a span of eigenspaces of l. Another isolating neighborhood of Sλµ is then
(B(3R/2) ∩ Vλµ ) × D, where D is a small closed ball in V̄ centered at the origin.
The flow ϕ̃µλ is then homotopic to the product of ϕµλ and a flow ψ on V̄ which
is generated by a vector field that is identical to l on D. From the definition of
the Conley index it is easy to see that IS 1 (ψ, {0}) ∼
= IS 1 (ψ0 , {0}). Here ψ0 is
Geometry & Topology, Volume 7 (2003)
911
Seiberg–Witten–Floer stable homotopy types
the linear flow generated by l on V̄ , and the corresponding equivariant Conley
index can be computed using Property 2 of the Conley index: it equals (Vλλ0 )+ .
By the same Property 2 we obtain
0
µ
µ
µ
I µ0 ∼
= IS 1 (ϕ̃ , S ) ∼
= (V λ0 )+ ∧ I .
λ
This implies that
0
Σ−Vλ Iλµ
and
λ
λ
−Vλ00 µ0
Σ
Iλ0
λ
λ
are canonically isomorphic.
Next we study what happens when we vary the metric on Y. We start by ex0
hibiting an isomorphism between the objects (Σ−Vλ Iλµ , 0, n(Y, c, g)) constructed
for two metrics g0 , g1 on Y sufficiently close to each other. Consider a smooth
homotopy (gt )0≤t≤1 between the two metrics, which is constant near t = 0.
We will use the subscript t to describe that the metric in which each object
is constructed is gt . Assuming all the gt are very close to each other, we can
arrange so that:
- there exist R, −λ∗ , µ∗ large enough and independent of t so that Proposition 3
is true for all metrics gt and for all values λ ≤ λ∗ , µ ≥ µ∗ ;
- there exist some λ < λ∗ and µ > µ∗ such that neither λ nor µ is an eigenvalue for any lt . Hence the spaces (Vλµ )t have the same dimension for all t, so
they make up a vector bundle over [0, 1]. Via a linear isomorphism that varies
continuously in t we can identify all (Vλµ )t as being the same space Vλµ ;
- for any t1 , t2 ∈ [0, 1] we have B(R)t1 ⊂ B(2R)t2 . Here we already think of the
balls as subsets of the same space Vλµ .
Then
\
B(2R)t
t∈[0,1]
is a compact isolating neighborhood for Sλµ in any metric gt with the flow (ϕµλ )t
on Vλµ . Note that (ϕµλ )t varies continuously in t. By Property 3 of the Conley
index,
µ
(Iλµ )0 ∼
= (Iλ )1 .
The difference nλ,0 − nλ,1 is the number of eigenvalue lines of −6 ∂ t , t ∈ [0, 1]
that cross the − line, counted with sign, ie, the spectral flow SF (−6 ∂ t ) as
defined in [3]. Atiyah, Patodi and Singer prove that it equals the index of the
operator ∂/∂t + 6 ∂ t on Ŷ = [0, 1] × Y with the metric gt on the slice t × Y and
with the vector ∂/∂t always of unit length.
Choose a 4–manifold X0 as in the previous section, with a neighborhood of the
boundary isometric to R+ × Y. We can glue Ŷ to the end of X0 to obtain a
manifold X1 diffeomorphic to X0 . Then
indC (D + ) = indC (D + ) + SF (−6 ∂ t )
Â,1
Geometry & Topology, Volume 7 (2003)
Â,0
912
Ciprian Manolescu
by excision. From the formula (6) and using the fact that c1 (L), χ and σ do
not depend on the metric we get
n(Y, c, g1 ) − n(Y, c, g0 ) = SF (−6 ∂ t ) = nλ,0 − nλ,1 .
It follows that (Vλ0 )0 ⊕ Cn(Y,c,g0) ∼
= (Vλ0 )1 ⊕ Cn(Y,c,g1) because the ∗d operator
has no spectral flow (for any metric its kernel is zero since H1 (Y ; R) = 0).
The orientation class of this isomorphism is canonical, because complex vector
spaces carry canonical orientations.
Thus we have constructed an isomorphism between the objects (Σ−Vλ Iλµ , 0, n(Y,
c, g)) for two different metrics close to each other. Since the space of metrics
M et is path connected (in fact contractible), we can compose such isomorphisms
and reach any metric from any other one.
0
In order to have an object in C well-defined up to canonical isomorphism, we
need to make sure that the isomorphisms obtained by going from one metric to
another along different paths are identical. Because M et is contractible, this
reduces to proving that when we go around a small loop in M et the construction
0
above induces the identity morphism on (Σ−Vλ Iλµ , 0, n(Y, c, g)). Such a small
loop bounds a disc D in M et, and we can find µ and λ so that they are not in
the spectrum of ∗d ⊕ 6 ∂ for any metric in D. Then the vector spaces Vλµ form
a vector bundle over D, which implies that they can all be identified with one
vector space, on which the Conley indices for different metrics are the same up
to canonical isomorphism. The vector spaces Vλ0 ⊕ Cn(Y,c,g) are also related to
each other by canonical isomorphisms in the homotopy category. Hence going
around the loop must give back the identity morphism in C.
A similar homotopy argument proves independence of the choice of R in Proposition 3. Thus SWF(Y, c) must depend only on Y and on its spinc structure,
up to canonical isomorphism in the category C.
8
The irreducible Seiberg–Witten–Floer invariants
In this section we construct a decomposition of the Seiberg–Witten–Floer invariant into its reducible and irreducible parts. This decomposition only exists
provided that the reducible θ is an isolated critical point of the Chern–Simons–
Dirac functional. To make sure that this condition is satisfied, we need to
depart here from our nonperturbative approach to Seiberg–Witten theory. We
introduce the perturbed Seiberg–Witten equations on Y :
∗(da − dν) + τ (φ, φ) = 0, 6 ∂ a φ = 0,
Geometry & Topology, Volume 7 (2003)
(9)
Seiberg–Witten–Floer stable homotopy types
913
where ν is a fixed L2k+1 imaginary 1-form on Y such that d∗ ν = 0.
In general, the solutions to (9) are the critical points of the perturbed Chern–
Simons–Dirac functional:
Z
1
CSDν (a, φ) = CSD(a, φ) +
a ∧ dν.
2 Y
Our compactness results (Proposition 1 and Proposition 3) are still true for the
perturbed Seiberg–Witten trajectories and their approximations in the finite
dimensional subspaces. The only difference consists in replacing the compact
map c with cν = c − ∗dν. There is still a unique reducible solution to the
equation (l + cν )(a, φ) = 0, namely θν = (ν, 0). Homotopy arguments similar
to those in the proof of Theorem 1 show that the SWF invariant obtained
from the perturbed Seiberg–Witten trajectories (in the same way as before) is
isomorphic to SWF(Y, c).
The advantage of working with the perturbed equations is that we can assume
any nice properties which are satisfied for generic ν. The conditions that are
needed for our discussion are pretty mild:
Definition 2 A perturbation ν ∈ L2k+1 (iΩ1 (Y )) is called good if ker 6 ∂ ν = 0
and there exists > 0 such that there are no critical points x of CSDν with
CSDν (x) ∈ (0, ).
Lemma 3 There is a Baire set of perturbations ν which are good.
Proof Proposition 3 in [12] states that there is a Baire set of forms ν for which
all the critical points of CSDν are nondegenerate. Nondegenerate critical points
are isolated. Since their moduli space is compact, we deduce that it is finite, so
there exists as required in Definition 2. Furthermore, the condition ker 6 ∂ ν = 0
is equivalent to the fact that the reducible θν is nondegenerate.
Let us choose a good perturbation ν, and let us look at the finite dimensional
approximation in the space Vλµ . For large µ and −λ, it is easy to see that
we must have ker(6 ∂ pµ ν ) = 0. This implies that θpµ ν , the reducible solution to
λ
λ
l + pµλ cν = 0, is an isolated critical point of CSDν |V µ . Note that CSDν (θpµ ν ) =
λ
λ
0, and there are no critical points x of CSDν |V µ with CSDν (x) ∈ (0, /2).
λ
Thus, in addition to S = Sλµ , we can construct four other interesting isolated
invariant sets for the flow ϕµλ :
Geometry & Topology, Volume 7 (2003)
914
Ciprian Manolescu
irr =the set of critical points x of CSD | µ with CSD (x) > 0, to• S>0
ν Vλ
ν
gether with all the trajectories between them; when it becomes necessary
irr )µ ;
to indicate the dependence on cutoffs, we will write (S>0
λ
irr = same as above, but with CSD (x) ≤ 0 and requiring x to be
• S≤0
ν
irreducible;
• S≤0 = same as above, with CSDν (x) ≤ 0 but allowing x to be the
reducible;
• finally, Θ = {θpµ ν }.
λ
Since every trajectory contained in S must end up in a critical point and CSDν
is decreasing along trajectories, S≤0 must be an attractor in S. Its dual repeller
irr , so by Proposition 4 we have a coexact sequence (omitting the flow ϕµ
is S>0
λ
and the group S 1 from the notation):
irr
I(S≤0 ) → I(S) → I(S>0
) → ΣI(S≤0 ) → · · ·
(10)
irr , Θ) is an attractor-repeller pair in S , so there is another coSimilarly, (S≤0
≤0
exact sequence:
irr
irr
I(S≤0
) → I(S≤0 ) → I(Θ) → ΣI(S≤0
) → ···
(11)
These two sequences give a decomposition of I(S) into several pieces which
are easier to understand. Indeed, I(Θ) ∼
= (Vλ0 )+ . Also, the intersection of S
µ
irr and S irr
with the fixed point set of Vλ is simply Θ. This implies that S>0
≤0
irr ) and I(S irr ) are
have neighborhoods in which the action of S 1 is free, so I(S>0
≤0
irr )µ the quotient of
S 1 –free as well (apart from the basepoint). Denote by (I>0
λ
irr
1
I(S>0 ) \ {∗} by the action of S .
Let us now rewrite these constructions to get something independent of the
cutoffs. Just like we did in the construction of SWF, we can consider the
following object of C :
−Vλ
irr µ
SWFirr
I(S>0
)λ
>0 (Y, c, g, ν) = Σ
0
and prove that it is independent of µ and λ (but not on the metric!) up to
canonical isomorphism. Similarly we get invariants SWFirr
≤0 and SWF≤0 . The
coexact sequences (10) and (11) give rise to exact triangles in the category C
(in the terminology of [21]):
SWF≤0 → (SWF, 0, n(Y, c, g)) → SWFirr
>0 → Σ(SWF≤0 ) → · · ·
0
irr
SWFirr
≤0 → SWF≤0 → S → Σ(SWF≤0 ) → · · ·
Furthermore, we could also consider the object
−Vλ irr µ
swfirr
(I>0 )λ
>0 (Y, c, g, ν) = Σ
0
Geometry & Topology, Volume 7 (2003)
Seiberg–Witten–Floer stable homotopy types
915
which lives in the nonequivariant graded suspension category (see [21]). This
1 action. It is indepenis basically the “quotient” of SWFirr
>0 under the S
dent of λ and µ, but not of the metric and perturbation. We could call it
the (nonequivariant) positive irreducible Seiberg–Witten–Floer stable homotopy
type of (Y, c, g, ν). Similarly we can define another metric-dependent invariant
swfirr
≤0 .
Remark If the flows ϕµλ satisfy the Morse-Smale condition, then the homology
of swfirr
swfirr
>0 (resp.
≤0 ) coincides with the Morse homology computed from
the irreducible critical points with CSDν > 0 (resp. CSDν ≤ 0) and the
trajectories between them. But we could also consider all the irreducible critical
points and compute a Morse homology SW HF (Y, c, g, ν), which is the usual
irreducible Seiberg–Witten–Floer homology (see [16], [20]). We expect a long
exact sequence:
irr
irr
· · · → H∗ (swfirr
≤0 ) → SW HF∗ → H∗ (swf>0 ) → H∗−1 (swf≤0 ) → · · ·
(12)
However, it is important to note that (12) does not come from an exact triangle
and, in fact, there is no natural stable homotopy invariant whose homology is
SW HF. The reason is that the interaction of the reducible with the trajectories
between irreducibles can be ignored in homology (it is a substratum of higher
codimension than the relevant one), but it cannot be ignored in homotopy.
9
Four-manifolds with boundary
In this section we prove Theorem 2. Let X be a compact oriented 4–manifold
with boundary Y. As in section 6, we let X have a metric such that a neighborhood of its boundary is isometric to [0, 1] × Y, with ∂X = {1} × Y. Assume
that X has a spinc structure ĉ which extends c. Let W + , W − be the two
spinor bundles, W = W + ⊕ W − , and L̂ the determinant line bundle. (We shall
often put a hat over the four-dimensional objects.) We also suppose that X
is homology oriented, which means that we are given orientations on H 1 (X; R)
2 (X; R).
and H+
Our goal is to obtain a morphism Ψ between the Thom space of a bundle over
the Picard torus P ic0 (X) and the stable homotopy invariant SWF(Y, c), with
a possible shift in degree. We construct a representative for this morphism as
the finite dimensional approximation of the Seiberg–Witten map for X.
Let Â0 be a fixed spinc connection on W. Then every other spinc connection on
W can be written Â0 + â, â ∈ iΩ1 (X). There is a corresponding Dirac operator
DÂ0 +â = DÂ0 + ρ̂(â),
Geometry & Topology, Volume 7 (2003)
916
Ciprian Manolescu
where ρ̂ denotes Clifford multiplication on the four-dimensional spinors. Let
ˆ An appropriate
C be the space of spinc connections of the form Â0 + ker d.
Coulomb gauge condition for the forms on X is â ∈ Im(dˆ∗ ), â|∂X (ν) = 0,
where ν is the unit normal to the boundary and dˆ∗ is the four-dimensional d∗
operator. Denote by Ω1g (X) the space of such forms. Then, for each µ ≥ 0 we
have a Seiberg–Witten map
g µ : C × (iΩ1 (X) ⊕ Γ(W + )) → C × (iΩ2 (X) ⊕ Γ(W − ) ⊕ V µ )
SW
g
+
(Â, â, φ̂) → (Â, F +
Â+â
− σ(φ̂, φ̂), DÂ+â (φ̂), pµ Πi∗ (â, φ̂))
Here i∗ is the restriction to Y , Π denotes Coulomb projection (the nonlinear
µ
map defined in section 3), and pµ is the orthogonal projection to V µ = V−∞
.
µ
g
Note that SW is equivariant under the action of the based gauge group
Gˆ0 = Map0 (X, S 1 ); this acts on connections in the usual way, on spinors by mulg µ /Gˆ0 is an S 1 –equivariant,
tiplication, and on forms trivially. The quotient SW
fiber preserving map over the Picard torus
P ic0 (X) = H 1 (X; R)/H 1 (X; Z) = C/Gˆ0 .
Let us study the restriction of this map to a fiber (corresponding to a fixed
 ∈ C ):
SW µ : iΩ1g (X) ⊕ Γ(W + ) → iΩ2+ (X) ⊕ Γ(W − ) ⊕ V µ .
Note that SW µ depends on µ only through its V µ –valued direct summand iµ ;
we write SW µ = sw ⊕ iµ . The reason for introducing the cut-off µ is that we
want the linearization of the Seiberg–Witten map to be Fredholm.
Let us decompose SW µ into its linear and nonlinear parts:
Lµ = d+ , DÂ , pµ (prker d∗ i∗ ) ; C µ = SW µ − Lµ .
Here prker d∗ is a shorthand for (prker d∗ , id) acting on the 1-forms and spinors
on Y, respectively.
As in [26], we need to introduce fractionary Sobolev norms. For the following
result we refer to [2] and [26]:
Proposition 5 The linear map
Lµ : L2k+3/2 iΩ1g (X) ⊕ Γ(W + ) → L2k+1/2 iΩ2+ (X) ⊕ Γ(W − ) ⊕ L2k+1 (V µ )
is Fredholm and has index
µ
2indC (D + ) − b+
2 (X) − dim V0 .
Â
Geometry & Topology, Volume 7 (2003)
917
Seiberg–Witten–Floer stable homotopy types
Here indC (D + ) is the index of the operator D + acting on the positive spinors
Â
Â
φ̂ with spectral boundary condition p0 i∗ (φ̂) = 0.
Equivalently, there is a uniform bound for all x̂ ∈ iΩ1g (X) ⊕ Γ(W + ):
kx̂kL2
k+3/2
≤ C(µ)· k(d+ ⊕ DÂ )x̂kL2
k+1/2
+ kpµ prker d∗ i∗ (x̂)kL2
k+1
+ kx̂kL2 ,
for some constant C(µ) > 0.
The nonlinear part is:
C µ : L2k+3/2 iΩ1g (X) ⊕ Γ(W + ) →
L2k+1/2 iΩ2+ (X) ⊕ Γ(W − ) ⊕ H 1 (X; R) ⊕ L2k+1 (V µ )
C µ (â, φ̂) = F + − σ(φ̂, φ̂), ρ̂(â)φ̂, 0, pµ (Π − (prker d∗ , id))i∗ (â, φ̂) .
Â
Just like in three dimensions, the first three terms are either constant or quadratic in the variables so they define compact maps between the respective
Sobolev spaces L2k+3/2 and L2k+1/2 . The last term is not compact. However, as
will be seen in the proof, it does not pose problems to doing finite dimensional
approximation. The use of this technique will lead us to the definition of Ψ,
the invariant of 4–manifolds with boundary mentioned in the introduction.
Let Un be any sequence of finite dimensional subspaces of L2k+1/2 (iΩ2+ (X) ⊕
Γ(W − )) such that prUn → 1 pointwise. For each λ < 0, let Un0 = (Lµ )−1 (Un ×
Vλµ ) and consider the map
prUn ×V µ SW µ = Lµ + prUn ×V µ C µ : Un0 → Un × Vλµ .
λ
λ
It is easy to see that for all n sufficiently large, Lµ restricted to Un0 (with values
in Un × Vλµ ) has the same index as Lµ . Indeed, the kernel is the same, while
the cokernel has the same dimension provided that Un × Vλµ is transversal to
the image of Lµ . Since prUn → 1 pointwise when n → ∞, it suffices to show
that Vλµ is transversal to the image of pµ (prker d∗ i∗ ) in V µ . But it is easy to see
that pµ (prker d∗ i∗ ) is surjective.
We have obtained a map between finite dimensional spaces, and we seek to get
from it an element in a stable homotopy group of Iλµ in the form of a map
between (Un0 )+ and (Un )+ ∧ Iλµ .
This can be done as follows. Choose a sequence n → 0, and denote by
B(Un , n ) and S(Un , n ) the closed ball and the sphere of radius n in Un (with
the L2k+1/2 norm), respectively. Let K̃ be the preimage of B(Un , n ) × Vλµ under the map Lµ , and let K̃1 and K̃2 be the intersections of K̃ with B(Un0 , R0 )
Geometry & Topology, Volume 7 (2003)
918
Ciprian Manolescu
and S(Un0 , R0 ), respectively. Here R0 is a constant to be defined later, and Un0
has the L2k+3/2 norm. Finally, let K1 , K2 be the images of K̃1 and K̃2 under
the composition of SW µ with projection to the factor Vλµ . Assume that there
exists an index pair (N, L) for Sλµ such that K1 ⊂ N and K2 ⊂ L. Then we
could define the pointed map we were looking for:
B(Un0 , R0 )/S(Un0 , R0 ) → (B(Un , n ) × N )/(B(Un , n ) × L ∪ S(Un , n ) × N ),
by applying prUn ×V µ ◦ SW µ to the elements of K̃ and sending everything else
λ
to the basepoint. Equivalently, via a homotopy equivalence we would get a
map:
Ψn,µ,λ, : (Un0 )+ → (Un )+ ∧ Iλµ .
Of course, for this to be true we need to prove:
Proposition 6 For µ, −λ sufficiently large and n sufficiently large compared
to µ and −λ, there exists an index pair (N, L) for Sλµ such that K1 ⊂ N and
K2 ⊂ L.
Let us first state an auxiliary result that will be needed. The proof follows from
the same argument as the proof of Proposition 3, so we omit it.
Lemma 4 Let t0 ∈ R. Suppose µn , −λn → ∞, and we have approximate
Seiberg–Witten half-trajectories xn : [t0 , ∞) → L2k+1 (Vλµnn ) such that xn (t) ∈
B(2R) for all t ∈ [t0 , ∞). Then xn (t) ∈ B(R) for any t > t0 and for any n
sufficiently large. Also, for any s > t0 , a subsequence of xn (t) converges to
some x(t) in C m norm, uniformly in t for t ∈ [s, ∞) and for any m > 0.
Proof of Proposition 6 We choose an isolating neighborhood for Sλµ to be
B(2R) ∩ Vλµ . Here R, the constant in Proposition 3, is chosen to be large
enough so that B(R) contains the image under iµ of the ball of radius R0 in
L2k+3/2 (iΩ1g (X) ⊕ Γ(W + )). By virtue of Theorem 4, all we need to show is that
K1 and K2 satisfy conditions (i) and (ii) in its hypothesis.
Step 1 Assume that there exist sequences µn , −λn → ∞ and a subsequence
of Un (denoted still Un for simplicity) such that the corresponding K1 do not
satisfy (i) for any n. Then we can find (ân , φ̂n ) ∈ B(Un0 , R0 ) and tn ≥ 0 such
that
prUn ×V µn ◦ SW µn (ân , φ̂n ) = (un , xn )
λn
with
kun kL2
k+1/2
≤ n ; (ϕµλnn )[0,∞) (xn ) ⊂ B(2R); (ϕµλnn )tn (xn ) ∈ ∂B(2R).
Geometry & Topology, Volume 7 (2003)
919
Seiberg–Witten–Floer stable homotopy types
We distinguish two cases: when tn → ∞ and when tn has a convergent subsequence. In the first case, let
yn : R → L2k+1 (Vλµnn )
be the trajectory of ϕµλnn such that yn (−tn ) = xn . Then, because of our hypotheses, kyn (0)kL2
= 2R and yn (t) ∈ B(2R) for all t ∈ [−tn , ∞). Since
k+1
tn → ∞, by Lemma 4 we have that yn (0) ∈ B(R) for n sufficiently large. This
is a contradiction.
In the second case, by passing to a subsequence we can assume that tn → t∗ ≥ 0.
We use a different normalization:
yn : [0, ∞) → L2k+1 (Vλµnn )
is the trajectory of ϕµλnn such that yn (0) = xn . Then kyn (tn )kL2
k+1
= 2R and
yn (t) ∈ B(2R) for all t ≥ 0. By the Arzela–Ascoli Theorem we know that yn
converges to some y : [0, ∞) → V in L2k norm, uniformly on compact sets
of t ∈ [0, ∞). This y must be the Coulomb projection of a Seiberg–Witten
trajectory.
Let zn = yn − y. From Lemma 4 we know that the convergence zn → 0 can be
taken to be in C ∞ , but only over compact subsets of t ∈ (0, ∞). However, we
can get something stronger than L2k for t = 0 as well. Since l is self-adjoint,
there is a well-defined compact operator el : L2k+1 (V 0 ) → L2k+1 (V 0 ). We have
the estimate:
Z 1
∂ tl 0
kp0 zn (0) − el p0 zn (1)kL2
= k
(e p zn (t))dtkL2
k+1
k+1
∂t
0
Z 1
∂
≤
ketl p0
zn (t) + lzn (t) kL2 dt
k+1
∂t
0
But since yn and y are trajectories of the respective flows, if we denote πn =
pµλnn , π n = 1 − πn we have
∂
zn (t) + lzn (t) = c(y(t)) − πn c(yn (t)),
∂t
so that
Z
kp zn (0) − e p zn (1)kL2
0
l 0
k+1
1
≤
0
ketl p0 πn c(y(t)) − c(yn (t)) kL2 dt+
k+1
Z
+
0
Geometry & Topology, Volume 7 (2003)
1
ketl p0 (π n c(y(t)))kL2
k+1
dt.
(13)
920
Ciprian Manolescu
Fix R > 0.R We break each of the two integrals on the right hand side of (13)
1
into 0 + . Recall that yn (t) live in B(2R). This must also be true for y(t)
because of the weak convergence yn (t) → y(t) in L2k+1 . Since etl p0 and c are
continuous maps from L2k+1 to L2k+1 , there is a bound:
Z Z tl 0
ke p πn c(y(t)) − c(yn (t)) kL2 dt +
ketl p0 (π n c(y(t)))kL2 dt ≤ C1 · ,
k+1
0
k+1
0
(14)
where C1 is a constant independent of .
On the other hand, on the interval [, 1] we have ketl p0 k ≤ kel p0 k and el p0 is
a compact map from L2k+1 to L2k+1 . We get
Z 1
Z 1
tl 0 n
ke p (π c(y(t)))kL2 dt ≤
kπ n el p0 c(y(t))kL2 dt.
k+1
k+1
In addition, el p0 c(y(t)) live inside a compact set of L2k+1 (V ) and we know that
π n → 0 uniformly on such sets. Therefore,
Z 1
ketl p0 (π n c(y(t)))kL2 dt → 0.
(15)
k+1
Similarly, using the fact that yn (t) → y(t) in L2k+1 (V ) uniformly in t for
t ∈ [, 1], we get:
Z 1
ketl p0 πn c(y(t)) − c(yn (t)) kL2 dt → 0.
(16)
k+1
Putting (13), (14), (15), and (16) together and letting → 0 we obtain:
kp0 zn (0) − el p0 zn (1)kL2
k+1
→ 0.
Since zn (1) → 0 in L2k+1 , the same must be true for p0 zn (0). Recall that
zn (0) = xn is the boundary value of an approximate Seiberg–Witten solution
on X :
prUn ×V µn ◦ SW µn (ân , φ̂n ) = (un , xn ),
λn
with kun kL2
k+1/2
≤ n . Equivalently,
Lµn + prUn ×V µn ◦ C µn (ân , φ̂n ) = (un , xn ).
λn
Since x̂n = (ân , φ̂n ) are uniformly bounded in L2k+3/2 norm, after passing to
a subsequence we can assume that they converge to some x̂ = (â, φ̂) weakly
Geometry & Topology, Volume 7 (2003)
921
Seiberg–Witten–Floer stable homotopy types
in L2k+3/2 . Changing everything by a gauge, we can assume without loss of
generality that i∗ (â) ∈ ker d∗ . Now Proposition 5 says that:
kx̂n − x̂kL2
k+3/2
≤ C(0)· k(d+ ⊕ DÂ )(x̂n − x̂)kL2
k+1/2
+ kp0 prker d∗ i∗ (x̂n − x̂)kL2
k+1
+ kx̂n − x̂kL2 . (17)
We already know that the last term on the right hand side goes to 0 as n → ∞.
Let us discuss the first term. First, it is worth seeing that sw(x̂) = 0. Let sw =
ˆl + ĉ be the decomposition of sw into its linear and compact parts; ˆl and ĉ are
direct summands of Lµ and C µ , respectively. We have prUn sw(x̂n ) = un → 0
in L2k+1/2 (because n → 0 by construction), and
sw(x̂) − prUn sw(x̂n ) = ˆl(x̂ − x̂n ) + prUn (ĉ(x̂) − ĉ(x̂n )) + (1 − prUn )ĉ(x̂).
Using the fact that x̂n → x̂ weakly in L2k+3/2 we get that each term on the
right hand side converges to 0 weakly in L2k+1/2 . Hence sw(x̂) = 0. Now the
first term on the right hand side of (17) is
ˆl(x̂n − x̂) = un + pr (ĉ(x̂) − ĉ(x̂n )) + (1 − pr )ĉ(x̂).
Un
Un
It is easy to see that this converges to 0 in L2k+1/2 norm. We are using here
the fact that prUn → 1 uniformly on compact sets.
Similarly one can show that the second term in (17) converges to 0. We already
know that p0 pµλnn Πi∗ (x̂n ) = p0 xn converges to p0 x. This was proved starting
from the boundedness of the yn on the cylinder on the right. In the same way,
using the boundedness of the x̂n on the manifold X on the left (which has a
cylindrical end), it follows that p0 xn → p0 x in L2k+1 . Thus, xn → x in L2k+1 . Let
i∗ (x̂n ) = (an + dbn , φn ) with an ∈ ker d∗ . We know that xn = pµλnn (an , eibn φn )
converges. Also, x̂n → x̂ weakly in L2k+3/2 , hence strongly in L2k+1/2 . This
implies that dbn → 0 in L2k and bn → 0 in L2k+1 . Since p0 prkerd∗ i∗ x̂n =
p0 (an , φn ) ∈ Vλ0n , they must converge in L2k+1 just like the xn . We are using
the Sobolev multiplication L2k+1 × L2k+1 → L2k+1 .
Putting all of these together, we conclude that the expression in (17) converges
to 0. Thus x̂n → x̂ in L2k+3/2 . We also know that sw(x̂) = 0. In addition, since
i∗ (x̂n ) → i∗ (x̂) in L2k+1 and using pµλnn Πi∗ (x̂n ) = xn we get that xn = yn (0) →
y(0) in L2k+1 . This implies that Πi∗ (x̂) = y(0).
Now it is easy to reach a contradiction: by a gauge transformation û of x̂
on X we can obtain a solution of the Seiberg–Witten equations on X with
Πi∗ (û· x̂) = y(0). Recall that y(0) was the starting point of y : [0, ∞) → B(2R),
Geometry & Topology, Volume 7 (2003)
922
Ciprian Manolescu
the Coulomb projection of a Seiberg–Witten half-trajectory of finite type. By
gluing this half-trajectory to û · x̂ we get a C 0 monopole on the complete
manifold X ∪ (R+ × Y ). From Proposition 2 we know that there are “universal”
bounds on the C m norms of the monopole (in some gauge) restricted to any
compact set, for any m. These bounds are “universal” in the sense that they
depend only on the metric on X. In particular, since Coulomb projection is
continuous, we obtain such a bound B on the L2k+1 norm of y(t) for all t.
Recall that yn (tn ) → y(t∗ ) because tn → t∗ , and that kyn (tn )kL2 = 2R.
k+1
When we chose the constant R, we were free to choose it as large as we wanted.
Provided that 2R > B, we get the desired contradiction.
Step 2 The proof is somewhat similar to that in Step 1.
Assume that there exist sequences µn , −λn → ∞ and a subsequence of Un
(denoted still Un for simplicity) such that the corresponding K2 do not satisfy
condition (ii) in Theorem 4 for any n. Then we can find x̂n ∈ S(Un0 , R0 ) such
that
prUn ×V µn ◦ SW µn (x̂n ) = (un , xn ),
λn
with
kun kL2
k+1/2
≤ n ; (ϕµλnn )[0,∞) (xn ) ⊂ B(2R).
Let yn : [0, ∞) → L2k+1 (Vλµnn ) be the half-trajectory of ϕµλnn starting at yn (0) =
xn . Repeating the argument in Step 1, after passing to a subsequence we can
assume that yn (t) converges to some y(t) in L2k (Vλµnn ), uniformly over compact
sets of t. Also, this convergence can be taken to be in C ∞ for t > 0, while for
t = 0 we get that p0 (yn (0) − y(0)) → 0 in L2k+1 . Observe that y is the Coulomb
projection of a Seiberg–Witten half-trajectory of finite type, which we denote
by y 0 . We can assume that y 0 (0) = y(0).
Then, just as in Step 1, we deduce that x̂n converges in L2k+3/2 to x̂, a solution
of the Seiberg–Witten equations on X with Πi∗ (x̂) = y(0). By gluing x̂ to y 0 we
obtain a C 0 monopole on X ∪ (R+ × Y ). By Proposition 2, this monopole must
be smooth in some gauge, and when restricted to compact sets its C m norms
must be bounded above by some constant which depend only on the metric on
X. Since the four-dimensional Coulomb projection from iΩ1 (X) ⊕ Γ(W + ) to
iΩ1g (X) ⊕ Γ(W + ) is continuous, we get a bound B 0 on the L2k+3/2 norm of x̂.
But x̂n → x̂ in L2k+3/2 and kx̂n kL2
= R0 . Provided that we have chosen the
k+3/2
constant R0 to be larger than B 0 , we obtain a contradiction.
Thus we have constructed some maps
Ψn,µ,λ, : (Un0 )+ → (Un )+ ∧ Iλµ
Geometry & Topology, Volume 7 (2003)
923
Seiberg–Witten–Floer stable homotopy types
for any µ, −λ sufficiently large and for all n sufficiently large compared to
µ and −λ. In other words, we get such maps from (U 0 )+ to U + ∧ Iλµ for
any µ and λ sufficiently large and for any finite dimensional subspace U ⊂
−
L2k+1/2 (iΩ+
2 (X) ⊕ Γ(W )) which contains a fixed subspace U0 (depending on
µ and λ).
For µ 0, the linear map Lµ is injective, because in the limit µ → ∞ there are
no nonzero solutions to an elliptic equation on X which vanish on the boundary.
For U × Vλµ transversal to coker Lµ and for U 0 = (Lµ )−1 (U × Vλµ ), we get a
natural identification:
U ⊕ Vλµ ∼
= U 0 ⊕ cokerLµ .
It is not hard to see that there is another natural identification:
µ
µ
∗
0
cokerLµ ∼
∗ i ) : ker L → V ).
= cokerL0 ⊕ coker(p (pr
0
Using the fact that
pµ0 (prker d∗ i∗ )
0
ker d
V0µ is
0
0
→
injective, we get:
µ
ker L ⊕ cokerLµ ∼
= cokerL ⊕ V0 .
:
ker L0
Consequently, the map
0 µ
(U 0 )+ → U + ∧ Iλµ ∼
= U + ∧ (Vλ0 )+ ∧ Σ−Vλ Iλ
is stably the same as a map:
(ker L0 )+ → (cokerL0 )+ ∧ Σ−Vλ Iλµ .
0
(18)
The real part of L0 is the (d+ , p0 i∗ ) operator restricted to Im(d∗ ). This has zero
2 (X; R). Using our homology orientation,
kernel, and cokernel isomorphic to H+
+
we can identify the latter with Rb2 (X) . The complex part of L0 is D + , which
Â
may have nontrivial kernel and cokernel. Assuming that all our constructions
have been done S 1 –equivariantly, (18) produces a stable equivariant morphism:
+
(ker D+ )+ → (cokerD + ⊕ Cn(Y,c,g) ⊕ Rb2 (X) )+ ∧ SWF(Y, c).
Â
Â
(19)
We can put these maps together for all classes [Â] ∈ P ic0 (X) as follows. We
started our construction from a bundle map between two Hilbert bundles over
the Picard torus P ic0 (X). Such bundles are trivial by Kuiper’s theorem, so
we can choose subbundles of the form U × P ic0 (X) when doing the finite
dimensional approximation. The maps (U 0 )+ → U + ∧ Iλµ can be grouped into
an S 1 –map from the Thom space of the vector bundle over P ic0 (X) with fibers
U 0 . In the process of stabilization, these U 0 –bundles differ from each other only
by taking direct sums with trivial bundles. In the end the collection of maps
(19) produces an S 1 –stable equivariant homotopy class:
Ψ ∈ {(T (Ind), b+
2 (X), 0), SWF(Y, c)}S 1 ,
Geometry & Topology, Volume 7 (2003)
924
Ciprian Manolescu
where T (Ind) is the Thom space of the virtual index bundle over P ic0 (X) of
the Dirac operator D+ , with a shift in complex degree by n(Y, c, g).
The class Ψ is independent of µ, λ, U and the other choices made in the construction, such as n and R0 . This can be seen using standard homotopy arguments analogous to those in the proof of Theorem 1. To interpolate between
different U, U 0 ⊃ U0 , it suffices to consider the case U ⊂ U 0 and use a linear homotopy tprU + (1 − t)prU 0 . Similar arguments show that Ψ does not depend on
the metric on X either, up to composition with canonical isomorphisms. Therefore we have constructed an invariant of X and its spinc structure, which we
denote Ψ(X, ĉ). This ends the proof of Theorem 2.
Remark 1 If we restrict Ψ to a single fiber of Ind we get an element in an
equivariant stable homotopy group:
1
S
ψ(X, ĉ) ∈ π̃−b,d
(SWF(Y, c)),
where b = b+
2 (X) and
c1 (L̂)2 − σ(X)
.
Â
8
(This is in fact given by the morphism (19) above.)
d = indC (D + ) − n(Y, c, g) =
1
Since π̃∗S is the universal equivariant homology theory, by composing with
the canonical map we obtain an invariant of X in h̃−b,d (SWF(Y, c)) for every
reduced equivariant homology theory h̃.
Remark 2 We can reinterpret the invariant ψ in terms of cobordisms. If Y1
and Y2 are 3–manifolds with b1 = 0, a cobordism between Y1 and Y2 is a 4–
manifold X with ∂X = Ȳ1 ∪ Y2 . Let us omit the spinc structures from notation
for simplicity. We have an invariant
S1
ψ(X) ∈ π̃−b,d
SWF(Ȳ1 ) ∧ SWF(Y2 ) = {(S 0 , b, −d), (SWF(Ȳ1 ) ∧ SWF(Y2 )}S 1 .
In [8], Cornea proves a duality theorem for the Conley indices of the forward
and reverse flows in a stably parallelizable manifold. This result (adapted to
the equivariant setting) shows that the spectra SWF(Y1 ) and SWF(Ȳ1 ) are
equivariantly Spanier-Whitehead dual to each other. According to [19], this
implies the equivalence:
{(S 0 , b, −d), (SWF(Ȳ1 ) ∧ SWF(Y2 )}S 1 ∼
= {SWF(Y1 ), (SWF(Y2 ), −b, d)}S 1 .
Therefore, a cobordism between Y1 and Y2 induces an equivariant stable homotopy class of S 1 −maps between SWF(Y1 ) and SWF(Y2 ), with a possible shift
in degree:
DX ∈ {SWF(Y1 ), (SWF(Y2 ), −b, d)}S 1 .
Geometry & Topology, Volume 7 (2003)
Seiberg–Witten–Floer stable homotopy types
10
925
Four-manifolds with negative definite intersection form
In [4], Bauer and Furuta give a proof of Donaldson’s theorem using the invariant
Ψ(X, ĉ) for closed 4–manifolds. Along the same lines we can use our invariant
to study 4–manifolds with boundary with negative definite intersection form.
The bound that we get is parallel to that obtained by Frøyshov in [12].
If Y is our 3–manifold with b1 (Y ) = 0 and spinc structure c, we denote by
s(Y, c) the largest s such that there exists an element
[f ] ∈ {(S 0 , 0, −s), SWF(Y, c)}S 1
which is represented by a pointed S 1 –map f whose restriction to the fixed
point set has degree 1. Then we set
s(Y ) = max s(Y, c).
c
The first step in making the invariant s(Y ) more explicit is the following lemma
(which also appears in [5]):
Lemma 5 Let f : (Rm ⊕ Cn+d)+ → (Rm ⊕ Cn )+ be an S 1 –equivariant map
such that the induced map on the fixed point sets has degree 1. Then d ≤ 0.
Proof Let fc be the complexification of the map f. Note that C ⊗R C =
V (1) ⊕ V (−1), where V (j) is the representation S 1 × C → C, (q, z) → q j z.
Using the equivariant K-theory mapping degree, tom Dieck proves in [9, II.5.15]
the formula:
1
d(fc ) = lim d(fcS )·tr λ−1 ([nV (1)⊕ nV (−1)]− [(n + d)V (1)⊕ (n + d)V (−1)])(q),
q→1
where q ∈ S 1 , d is the usual mapping degree, and λ−1 ([nV (1) ⊕ nV (−1)] −
[(n + d)V (1) ⊕ (n + d)V (−1)]) is the KS 1 –theoretic Euler class of fc ; in our
1
case its character evaluated at q equals (1 − q)−d (1 − q −1 )−d . Since d(fcS ) = 1,
the limit only exists in the case d ≤ 0.
Example Let us consider the case when Y is the Poincaré homology sphere P,
oriented as the link of the E8 singularity. There is a unique spinc structure c on
P, and P admits a metric g of positive scalar curvature. The only solution of
the Seiberg–Witten equations on P with the metric g is the reducible θ = (0, 0).
In addition, the Weitzenböck formula tells us that the operator 6 ∂ is injective,
Geometry & Topology, Volume 7 (2003)
926
Ciprian Manolescu
hence so is l. We can choose R as small as we want in Proposition 3. Taking
the L2k+1 norms, we get a bound
kpµλ c(v)k ≤ kc(v)k ≤ αkvk2
for all v ∈ V sufficiently close to 0. Also, if λ0 is the eigenvalue of l of smallest
absolute value, then
kl(v)k ≥ |λ0 | · kvk.
Putting the two inequalities together, we get that for R > 0 sufficiently small
and −λ, µ sufficiently large, the only zero of the map l + pµλ c in B(2R) is 0. It
0
0
follows that Sλµ = Inv B(2R)∩Vλµ = {0}. Its Conley index is (Rmλ )+ ∧(Cnλ )+ .
In [12], K. Frøyshov computed n(P, c, g) = −1, so that we can conclude:
SWF(P, c) = C+
up to isomorphism. We get that s(P ) = 1 as a simple consequence of Lemma 5.
Let us come back to the general case and try to obtain a bound on s(Y ). Recall
the notations from section 8. Choose a metric g and a good perturbation ν.
We seek to find s so that there is no element in {(S 0 , 0, −s − 1), SWF(Y, c)}
representable by a map which has degree 1 on the fixed point sets. Equivalently,
for µ and −λ sufficiently large, there should not be any S 1 –map f˜ of that kind
0
0
between (Rmλ ⊕ Cnλ +r+1 )+ and I(S), where r = s + n(Y, c, g). Assume that
r ≥ 0.
Suppose that there exists f˜ as above and denote N = m0 + 2n0 . Consider S 1 –
λ
λ
irr ), I(S irr ), I(S ), I(Θ) compatible
equivariant cell decompositions of I(S), I(S>0
≤0
≤0
with the coexact sequences (10) and (11) from section 8 in the sense that all
maps are cellular. We can assume that all the S 1 –cells of I(S) of cellular
irr ) and I(S irr ) are free. Also
dimension ≥ N and all the S 1 –cells of I(S>0
≤0
0
0 +r+1 +
m
n
1
note that (R λ ⊕ C λ
) has an S –cell structure with no equivariant
cells of cellular dimension greater than N + 2r + 1. Thus we can homotope
f˜ equivariantly relative to the fixed point set so that its image is contained in
the (N + 2r + 1)–skeleton of I(S). Assuming that there exists an S 1 –map
0
0
f : I(S)N +2r+1 → (Rmλ ⊕ Cnλ +r )+
whose restriction to the fixed point set has degree 1, by composing f˜ with f
we would get a contradiction with Lemma 5.
Therefore, our job is to construct the map f. Start with the inclusion:
0
0
0
0
I(Θ) ∼
= (Rmλ ⊕ Cnλ )+ ,→ (Rmλ ⊕ Cnλ +r )+
By composing with the second map in (11) and by restricting to the (N +2r+1)–
skeleton we obtain a map f0 defined on I(S≤0 )N +2r+1 . Let us look at the
Geometry & Topology, Volume 7 (2003)
Seiberg–Witten–Floer stable homotopy types
927
irr ) is S 1 –free, we could obtain the desired f once we
sequence (10). Since I(S>0
are able to extend f0 from I(S≤0 )N +2r+1 to I(S)N +2r+1 . This is an exercise
in equivariant obstruction theory. First, it is easy to see that we can always
extend f0 up to the (N +2r)–skeleton. Proposition II.3.15 in [9] tells us that the
extension to the (N +2r+1)–skeleton is possible if and only if the corresponding
obstruction
+2r+1
m0λ
n0λ +r +
irr µ
∼
γ̃r ∈ HN
I(S),
I(S
);
π
((R
⊕
C
)
)
)λ ; Z)
= H N +2r+1 ((I>0
≤0
N
+2r
1
S
vanishes. Here H denotes the Bredon cohomology theory from [9, Section II.3].
After stabilization, the obstruction γ̃r becomes an element
γr ∈ H 2r+1 (swfirr
>0 (Y, c, g, ν)).
Thus, we have obtained the following bound:
s(Y ) ≤ max inf −n(Y, c, g) + min{r ∈ Z+ |γr = 0} .
c
g,ν
(20)
We have now developed the tools necessary to study four-manifolds with negative definite intersection forms.
Proof of Theorem 3 A characteristic element c is one that satisfies c · x ≡
x · x mod 2 for all x ∈ H2 (X)/Torsion. Given such a c, there is a spinc
structure ĉ on X with c1 (L̂) = c.
Let d = (c2 − σ(X))/8. In section 9 we constructed an element:
ψ(X, ĉ) ∈ {(S 0 , 0, −d), SWF(Y, c)}S 1 .
The restriction to the fixed point set of one of the maps Ψn,µ,λ, which represents ψ(X, ĉ) is linear near 0 and has degree ±1 because b+
2 (X) = 0. Hence
d ≤ s(Y, c) ≤ s(Y ).
Together with the inequality (20), this completes the proof.
Corollary 1 (Donaldson) Let X be a closed, oriented, smooth four-manifold
with negative definite intersection form. Then its intersection form is diagonalizable.
Proof If we apply Theorem 3 for Y = ∅, we get b2 (X) + c2 ≤ 0 for all
characteristic vectors c. By a theorem of Elkies from [10], the only unimodular
forms with this property are the diagonal ones.
Geometry & Topology, Volume 7 (2003)
928
Ciprian Manolescu
Corollary 2 (Frøyshov) Let X be a smooth, compact, oriented 4–manifold
with boundary the Poincaré sphere P. If the intersection form of X is of the
form mh−1i ⊕ J with J even and negative definite, then J = 0 or J = −E8 .
Proof Since J is even, the vector c whose first m coordinates are 1 and the
rest are 0 is characteristic. We have c2 = −m and we have shown that s(P ) = 1.
Rather than applying Theorem 3, we use the bound d = b2 (X) + c2 ≤ 8s(P )
directly. This gives that rank(J) ≤ 8. But the only even, negative definite form
of rank at most 8 is −E8 .
A
Existence of index pairs
This appendix contains the proof of Theorem 4, which is an adaptation of the
argument given in [7], pages 46-48.
The proof is rather technical, so let us first provide the reader with some intuition. As a first guess for the index pair, we could take N to be the complement
in A of a small open neighborhood of ∂A ∩ A+ and L to be the complement
in N of a very small neighborhood of A+ . (This choice explains condition (ii)
in the statement of Theorem 3.) At this stage (N, L) satisfies conditions 1 and
2 in the definition of the index pair, but it may not satisy the relative positive
invariance condition. We try to correct this by enlarging N and L with the
help of the positive flow. More precisely, if B ⊂ A, we denote
P (B) = {x ∈ A : ∃y ∈ B, t ≥ 0 such that ϕ[0,t] (y) ⊂ A, x = ϕt (y)}.
We could replace N and L by P (N ) and P (L), respectively. (This explains
the condition (i) in the satement of Theorem 3, which can be rewritten P (K1 )∩
∂A ∩ A+ = ∅.) We have taken care of positive invariance, but a new problem
appears: P (N ) and P (L) may no longer be compact. Therefore, we need to
find conditions which guarantee their compactness:
Lemma 6 Let B be a compact subset of A which either contains A− or is
disjoint from A+ . Then P (B) is compact.
Proof Since P (B) ⊂ A and A is compact, it suffices to show that for any
xn ∈ P (B) with xn → x ∈ A, we have x ∈ P (B). Let xn be such a sequence,
xn = ϕtn (yn ), yn ∈ B so that ϕ[0,tn ] (yn ) ⊂ A. Since B is compact, by passing
to a subsequence we can assume that yn → y ∈ B. If tn have a convergent
Geometry & Topology, Volume 7 (2003)
Seiberg–Witten–Floer stable homotopy types
929
subsequence as well, say tnk → t ≥ 0, then by continuity ϕtnk (ynk ) → ϕt (y) = x
and ϕ[0,t] (y) ⊂ A. Thus x ∈ P (B), as desired.
If tn has no convergent subsequences, then tn → ∞. Given any m > 0, for n
sufficiently large tn > m, so ϕ[0,m] (yn ) ⊂ A. Letting n → ∞ and using the
comapctness of A we obtain ϕ[0,m] (y) ⊂ A. Since this is true for all m > 0,
we have y ∈ A+ . This takes care of the case A+ ∩ B = ∅, since we obtain a
contradiction. If A− ⊂ B, we reason differently: ϕ[0,tn ] (yn ) ⊂ A is equivalent
to ϕ[−tn ,0] (xn ) ⊂ A; letting n → ∞, we get ϕ(−∞,0] (x) ⊂ A, so x ∈ A− . Thus
x ∈ B ⊂ P (B), as desired.
Proof of Theorem 4 Choose C a small compact neighborhood of A+ ∩ ∂A
such that A− ∩ C = ∅. We claim that if we choose C sufficiently small, we have
P (K1 ) ∩ C = ∅. Indeed, if there were no such C, we could find xn ∈ P (K1 )
with xn → x ∈ A+ ∩ ∂A. Let xn = ϕtn (yn ) for some yn ∈ K1 such that
ϕ[0,tn ] (yn ) ⊂ A. By passing to a subsequence we can assume yn → y ∈ K1 . If
tn has a subsequence converging to some t ∈ [0, ∞), then by taking the limit
ϕ[0,t] (y) ⊂ A and ϕt (y) = x, which contradicts P (K1 ) ∩ A+ ∩ ∂A = ∅. If tn
has no such subsequence, then tn → ∞. Since ϕ[−tn ,0] (xn ) ⊂ A, by taking the
limit we get ϕ(−∞,0] (x) ⊂ A. Thus x ∈ A− . On the other hand x ∈ A+ ∩ ∂A,
which contradicts the fact that A+ ∩ A− = Inv A ⊂ int(A).
Let C be as above and let V be an open neighborhood of A+ such that cl(V \
C) ⊂ int(A). Since K2 ∩ A+ = ∅ and K2 is compact, by making V sufficiently
small we can assume that K2 ∩ V = ∅.
Let us show that there exists t∗ ≥ 0 such that ϕ[−t∗ ,0] (y) 6⊂ A for any y ∈ C. If
not, we could find yn ∈ C with ϕ[−n,0] (yn ) ⊂ A. Since C is compact, there is
a subsequence of yn which converges to some y ∈ C such that ϕ(−∞,0] (y) ⊂ A
or, equivalently, y ∈ A− . This contradicts the fact that A− and C are disjoint.
Let t∗ be as above. For each x ∈ A− , either ϕ[0,t∗ ] (x) ⊂ A− or there is
t(x) ∈ [0, t∗ ] so that ϕ[0,t(x)] (x) ∩ C = ∅ and ϕt(x) (x) 6∈ A. In the first case we
choose K(x) a compact neighborhood of x such that ϕ[0,t∗ ] (K(x)) ∩ C = ∅.
In the second case we choose K(x) to be a compact neighborhood of x with
ϕ[0,t(x)] (K(x)) ∩ C = ∅ and ϕt(x) (K(x)) ∩ A = ∅. Since A− is compact, it
is covered by a finite collection of the sets K(x). Let B 0 be their union and
let B = B 0 ∪ K1 . Then B is compact, and we can assume that it contains a
neighborhood of A− .
We choose the index pair to be
L = P (A \ V ); N = P (B) ∪ L.
Geometry & Topology, Volume 7 (2003)
930
Ciprian Manolescu
Clearly K1 ⊂ B ⊂ N and K2 ⊂ A \ V ⊂ L. It remains to show that (N, L) is
an index pair. First, since A \ V is compact and disjoint from A+ , by Lemma
6 above L is compact. Since A− ⊂ B, N = P (B ∪ (A \ V )) is compact as well.
We need to check the three conditions in the definition of an index pair. Condition 1 is equivalent to S ⊂ int(N \L) = int(N )\L. We have S ⊂ int(N ) because
S ⊂ A− and B ⊂ N contains a neighborhood of A− . We have S ∩ L = ∅ because if x ∈ S ⊂ A+ is of the form x = ϕt (y) for y ∈ A \ V, t ≥ 0 such that
ϕ[0,t] (y) ⊂ A, then y ∈ A+ , which contradicts A+ ⊂ V.
Condition 3 can be easily checked from the definitions: L is positively invariant
in A by construction, and this implies that it is positively invariant in N as
well.
Condition 2 requires more work. Let us first prove that P (B) ∩ C = ∅. We
have P (B) = P (B 0 ) ∪ P (K1 ) and we already know that P (K1 ) ∩ C = ∅. For
y 0 ∈ P (B 0 ), there exists y ∈ B 0 such that ϕ[0,t] (y) ⊂ A and ϕt (y) = y 0 . Recall
that we chose t∗ ≥ 0 so that ϕ[−t∗ ,0] (x) 6⊂ A for any x ∈ C.If t ≥ t∗ , this
implies y 0 6∈ C. If t < t∗ then, because y is in some K(x) for x ∈ A− , the fact
that ϕ[0,t] (y) ⊂ A implies again y 0 = ϕt (y) 6∈ C. Therefore P (B 0 ) ∩ C = ∅, so
P (B) ∩ C = ∅.
To prove that L is an exit set for N, pick x ∈ N \L and let τ = sup{t|ϕ[0,t] (x) ⊂
N \ L}. It suffices to show that ϕτ (x) ∈ L. Assume this is false; then ϕτ (x) ⊂
N \ L. Note that
N \ L ⊂ (A \ P (A \ V )) ⊂ V.
Also N \ L ⊂ P (B) ⊂ (A \ C), so N \ L is contained in V \ C ⊂ int(A).
It follows that for > 0 sufficiently small, ϕ[τ,τ +] (x) ⊂ A \ L. Since N is
positively invariant in A and ϕτ (x) ∈ N, we get ϕ[τ,τ +] (x) ⊂ N \ L. This
contradicts the definition of τ. Therefore, ϕτ (x) ∈ L.
We conclude that (N, L) is a genuine index pair, with K1 ⊂ N and K2 ⊂ L.
Geometry & Topology, Volume 7 (2003)
Seiberg–Witten–Floer stable homotopy types
931
References
[1] J F Adams, Prerequisites (on equivariant stable homotopy) for Carlsson’s lecture, from: “Algebraic topology, Aarhus 1982”, Springer–Verlag (1984) 483–532
[2] M F Atiyah, V K Patodi, I M Singer, Spectral assymetry and Riemannian
geometry: I, Math. Proc. Cambridge Philos. Soc. 77 (1975) 43–69
[3] M F Atiyah, V K Patodi, I M Singer, Spectral assymetry and Riemannian
geometry: III, Math. Proc. Cambridge Philos. Soc. 79 (1976) 71–99
[4] S Bauer, M Furuta, A stable cohomotopy refinement of Seiberg–Witten invariants: I, arXiv:math.DG/0204340
[5] S Bauer, A stable cohomotopy refinement of Seiberg–Witten invariants: II,
arXiv:math.DG/0204267
[6] R L Cohen, J D S Jones, G B Segal, Floer’s infinite-dimensional Morse theory and homotopy theory, from: “The Floer memorial volume”, Birkhäuser,
Basel (1995) 297–325
[7] C Conley, Isolated invariant sets and the Morse index, Amer. Math. Soc.,
Providence (1978)
[8] O Cornea, Homotopical dynamics: suspension and duality, Erg. Th. and Dyn.
Syst. 20 (2000) 379–391
[9] T tom Dieck, Transformation groups, de Gruyter studies in mathematics, 8;
de Gruyter (1987)
[10] N D Elkies, A characterization of the Zn lattice, Math. Res. Lett. 2 (1995)
321–326
[11] A Floer, A refinement of the Conley index and an application to the stability
of hyperbolic invariant sets, Ergod. Th. and Dynam. Syst. 7 (1987) 93–103
[12] K A Frøyshov, The Seiberg–Witten equations and four-manifolds with boundary, Math. Res. Lett. 3 (1996) 373–390
[13] M Furuta, Monopole equation and the 11/8-conjecture, Math. Res. Lett. 8
(2001) 279–291
[14] M Furuta, Stable homotopy version of Seiberg–Witten invariant, preprint, now
incorporated in [4]
[15] R E Gompf, A I Stipsicz, 4–manifolds and Kirby calculus, Graduate studies
in mathematics, vol.20, Amer. Math. Soc. 1999
[16] P B Kronheimer, Embedded surfaces and gauge theory in three and four dimensions, in Surveys in differential geometry, Vol. III (Cambridge, MA, 1996)
Int. Press (1998)
[17] P B Kronheimer, T S Mrowka, The genus of embedded surfaces in the projective plane, Math. Res. Lett. 1 (1994) 797–808
[18] P B Kronheimer, T S Mrowka, Floer homology for Seiberg–Witten
monopoles, in preparation
Geometry & Topology, Volume 7 (2003)
932
Ciprian Manolescu
[19] L G Lewis, J P May, M Steinberger, Equivariant stable homotopy theory,
Lecture notes in mathematics, vol. 1213, Springer–Verlag (1986)
[20] M Marcolli, B L Wang, Equivariant Seiberg–Witten Floer homology, Comm.
Anal. Geom. 9 (2001) 450–640
[21] H R Margolis, Spectra and the Steenrod algebra, North–Holland Mathematical
Library, vol. 29, North–Holland (1993)
[22] L I Nicolaescu, Notes on Seiberg–Witten theory, Graduate studies in mathematics, vol.28, Amer. Math. Soc. (2000)
[23] A M Pruszko, The Conley index for flows preserving generalized symmetries,
from: “Conley index theory”, Banach Center Publ. 47, Polish Acad. Sci., Warsaw (1999) 193–217
[24] D Salamon, Connected simple systems and the Conley index of isolated invariant sets, Trans. Amer. Math. Soc. 291 (1985) 1–41
[25] M Schwarz, Morse homology, Birkhäuser Verlag (1993)
[26] C H Taubes, L2 moduli spaces on 4–manifolds with cylindrical ends, Monographs in geometry and topology, vol.1, International Press (1993)
[27] G W Whitehead, Elements of homotopy theory, Springer–Verlag (1978)
Geometry & Topology, Volume 7 (2003)
Download